HARDWARE

What runs on a single GPU

A single GPU is the most common business setup: a server with a 16, 24, 48 or 80 GB card covers almost everything except the largest models. This is where live chat, image generation and real-time document processing become practical.

177 familiesUpdated 22 Sep 2026Calculate video memory for your model size

What this means in practice

16 GB cardMid-sized language models in compressed form, document understanding, image generation. The cheapest way in.
24 GB cardThe workhorse: a bigger model, or several tasks on one card. Enough for a chatbot and background processing together.
48-80 GB cardModels at full precision, long context, video generation. More expensive, but one card replaces a small cluster.

What runs on it

Image generation 18

Image generationRU2022–2026

Kandinsky

Sber (Kandinsky Lab) · Russia

Sber's Russian family of image and video generation models. Understands Russian-language prompts and Russian cultural context well; released under MIT. The new Kandinsky S2V 1.0 (Avatar, MIT) animates a person from a single frame to a speech recording: the whole figure moves, with expressions and gestures, in clips of up to a minute. Pro (HD) and Lite (SD, runs on a home graphics card) versions; it was trained on more than a million fragments in 95 languages.

  • Images from Russian-language descriptions
  • Short promo videos from text or a photo
  • Instruction-based image editing
Sizes
2B – 21B
Hardware
from: 1 GPU
Commercial use allowedDetails
Image generationGGUF2025–2026

Chroma

lodestones (independent developer) · not disclosed

A community model retrained from FLUX.1-schnell with a simplified architecture. No style censorship; popular as a base for fine-tuning. In October Chroma-H3 (lodestones) came out: an image fine-tune of the 33B MiniMax H3 video model with speed-up adapters for 4 and 8 steps; its license comes from MiniMax H3.

  • Base for fine-tuning your own styles
  • Artistic illustrations
  • Images for games
Sizes
4B – 33B
Hardware
from: 1 GPU
Commercial use with conditionsDetails
Image generation2026

Fibo 2.0

Bria AI · Israel

A family of Bria image models: Fibo 2.0 dev draws from text, Fibo 2.0 edit turbo edits finished images from a description. According to the authors' paper, the model was trained on very detailed structured descriptions so that it follows a long brief better. The weights are free only for non-commercial work: advertising and sales need a separate Bria license. There are ready files for ComfyUI, including ones compressed to int8. The languages of prompts are not stated on the card.

  • creating pictures from a detailed description for tests and prototypes
  • editing finished pictures with a text command
  • research and personal projects without commerce
Sizes
not stated (the bundle includes a Qwen3-VL 4B text encoder)
Hardware
from: 1 GPU
Non-commercial onlyDetails
Image generation2026

MOSS-Image

OpenMOSS · China

A text-to-image model. It takes an ordinary description, sets of tags and a JSON layout in which you can set where a caption stands and how objects are arranged. It renders text in English and Chinese well. Base and fine-tuned (SFT) versions exist, and it plugs into ComfyUI. It uses a text encoder based on Qwen3.5-9B and the FLUX.2 decoder. The community is already fine-tuning it: Kromoss (lodestones) removes the blurriness of the base version, is a drop-in replacement for SFT and comes with an 8-step speed-up LoRA.

  • banners and covers with captions
  • infographics and posters with a given layout
  • pictures for product cards
Sizes
about 5B
Hardware
from: 1 GPU
Commercial use allowedDetails
Image generationGGUF2024–2026

FLUX

Black Forest Labs · Germany

Image generation from the creators of Stable Diffusion. Renders text in images well and keeps the composition. The new FLUX.3 Image Klein (5.6B) targets ordinary graphics cards: up to 4K at once, regional editing, up to ten references; access to the weights is granted automatically after signing in on Hugging Face.

  • Images for product cards
  • Banners and covers
  • Photo editing by description (Kontext)
Sizes
4B – 32B
Hardware
from: 1 GPU
Commercial use with conditionsDetails
Image generation2026

Iris-3B

Sperid Labs · not stated

Generates images directly in pixels, without a compressed latent space. After fine-tuning, the same weights estimate depth from a single photo and restore and upscale images 4 times. Prompts are in English, the default size is 1024 by 1024.

  • generating images from a description
  • estimating scene depth from a single photo
  • restoring and upscaling old or compressed photos
Sizes
3B
Hardware
from: 1 GPU
Commercial use allowedDetails
Image generationGGUF2025–2026

Qwen-Image

Alibaba · China

Image generation and editing, including text in images. Earlier versions allow commercial use; the latest 2.1 is non-commercial only.

  • Infographics for product cards
  • Photo editing by text command
  • Ad creatives
Sizes
7B – 20B
Hardware
from: 1 GPU
Commercial use with conditionsDetails
Image generationGGUF2026

Ideogram 4

Ideogram · Canada

Open weights of the Ideogram model, known for precise typography. Under a non-commercial license: for business, suitable only for testing.

  • Testing text-in-image generation
  • Research and prototypes
Sizes
about 9B
Hardware
from: 1 GPU
Non-commercial onlyDetails
Image generationGGUF2026

Krea 2

Krea · USA

A 12B image model focused on realism without the glossy "AI look". The Turbo version produces a 2K image in a couple of seconds.

  • Realistic photos for advertising
  • High-resolution images
  • Style fine-tuning
Sizes
12B
Hardware
from: 1 GPU
Commercial use with conditionsDetails
Image generationGGUF2025–2026

HiDream

HiDream.ai · China

Open MIT-licensed image models: generation (I1), instruction-based editing (E1) and the unified O1-Image model that does both.

  • Image generation from descriptions
  • Editing images with words
  • Variations of product photos
Sizes
about 9B – 17B
Hardware
from: 1 GPU
Commercial use allowedDetails
Image generationGGUF2025–2026

LongCat-Image

Meituan · China

Meituan's 6B image generation and editing model. Renders Chinese text well; has a fast version for edits.

  • Image generation from descriptions
  • Instruction-based photo editing
  • Visuals for product cards
Sizes
6B
Hardware
from: 1 GPU
Commercial use allowedDetails
Image generation2026

GLM-Image

Zhipu AI (Z.ai) · China

A hybrid of a 9B language model and a 7B decoder. Strong at text-heavy images: posters, infographics, slides.

  • Posters and banners with text
  • Infographics
  • Illustrations for presentations
Sizes
9B + 7B
Hardware
from: 1 GPU
Commercial use allowedDetails

Video 18

Video2026

Hy-Nomas

Tencent Hunyuan · China

A "living world" video model: it builds the world segment by segment, and you steer it with text and a camera trajectory. At each step you can add an object, change the style or scene, describe an action or turn the camera, and the memory of the world carries over between segments. Output is 720 by 1280 at 24 frames per second, and a segment takes 8 steps. The standard configuration is meant for several graphics cards, but there is a variant for a single RTX 4090 with INT8 weights and part of the model offloaded to the CPU. For narrow areas such as robotics a light LoRA tuning can be attached. Together with the model came the Hy-Nomas-Bench set of 280 test cases; the technical report and tests of the model itself are promised later.

  • interactive video worlds and game-scene prototypes from a description
  • test videos for robotics with LoRA tuning
  • experiments with a controllable camera and scene consistency
Sizes
not stated (54 MMDiT blocks, a Qwen2.5-VL-7B text encoder)
Hardware
from: 1 GPU
Commercial use allowedDetails
Video2026

Prism

Fudan University, Tencent Hunyuan and Zhejiang University · China

Creates video together with sound: from a first frame and a description you get a clip with speech, effects and music. Native generation at 720p, 1080p and 2K is stated. In the third-party FreeVideo build (1280 by 720, 8.5 seconds) it runs on one RTX 30, 40 or 50 series card with 12 GB or more, but it will not be fast.

  • animating a picture into a clip with sound
  • ad and training videos with voice-over
  • prototypes of scenes for video and animation
Sizes
not stated on the card (based on MOVA-360p)
Hardware
from: 1 GPU
Commercial use allowedDetails
VideoGGUF2025–2026

MAGI

Sand AI · China

Video generated chunk by chunk in sequence, so a clip can be extended indefinitely. MAGI-2 produces video with sound.

  • Long videos with continuation
  • Video with sound
  • Animating images
Sizes
4.5B – 114B-A6B
Hardware
from: 1 GPU
Commercial use allowedDetails
Video2025–2026

SANA-Video

NVIDIA · USA

NVIDIA's lightweight, fast video model. Produces 720p clips on a single GPU; a 4-step version enables quick generation.

  • Quick clips for social media
  • Bulk video generation
  • Video from an image
Sizes
2B – 5B
Hardware
from: 1 GPU
Commercial use allowedDetails
Video2024–2026

LTX-Video / LTX-2

Lightricks · Israel

A fast video model; with LTX-2 it generates video with sound and speech in one go. Camera and pose control, lightweight versions available.

  • Ad videos with sound
  • Video from a product photo
  • Voiced scenes for social media
Sizes
2B – 22B
Hardware
from: 1 GPU
Commercial use with conditionsDetails
Video2025–2026

Wan

Alibaba · China

Text-to-video and image-to-video; the small version runs on a gaming GPU. After 2.2 only applied models are open: editing (VACE), audio-driven talking characters (S2V), dancing to music (Dancer).

  • Short promo videos
  • Animating product photos
  • Videos for social media
Sizes
1,3B – 14B
Hardware
from: 1 GPU
Commercial use allowedDetails
VideoGGUF2025–2026

SCAIL

Zhipu AI (Z.ai) and Tsinghua University · China

Animates a character from an image using motion from another video, including complex turns and multiple characters. SCAIL-2 works without an intermediate skeleton and can replace a character in a clip.

  • Transferring an actor's motion to a character
  • Replacing a character in a finished video
  • Animating mascots and illustrations
Sizes
14B
Hardware
from: 1 GPU
Commercial use allowedDetails
Video2025–2026

Matrix-Game

Skywork AI (Kunlun) · China

An interactive "world model": generates video of a game world in real time and responds to keyboard and mouse input. Version 3.0 keeps scene memory for minutes.

  • Game world prototypes without an engine
  • Interactive demos and simulations
  • Generating data to train agents
Sizes
1.8B – 17B
Hardware
from: 1 GPU
Commercial use allowedDetails
VideoGGUF2026

MOVA

OpenMOSS / MOSI · China

Generates video with sound in one pass: lip-synced speech, effects and ambience. A 32B-parameter MoE architecture, with 360p and 720p versions.

  • Short clips with speech and sound from a description
  • Ad scenes with dialogue
  • Video prototypes for storyboards
Sizes
32B-A18B
Hardware
from: 1 GPU
Commercial use allowedDetails
VideoGGUF2025–2026

SkyReels

Skywork AI (Kunlun Tech) · China

Video models for cinematic scenes with people. Can make videos of unlimited length, extend videos and create talking characters from audio.

  • Long videos with continuation
  • Video with one character from a reference
  • Talking avatar from a voice
Sizes
1.3B – 19B
Hardware
from: 1 GPU
Commercial use with conditionsDetails
VideoGGUF2024–2025

HunyuanVideo

Tencent · China

Tencent's video model, one of the first open ones on par with closed services. Version 1.5 is lighter (8.3B) and runs on consumer GPUs.

  • Video from a text script
  • Animating images
  • Base for fine-tuning your own video models
Sizes
8.3B – 13B
Hardware
from: 1 GPU
Commercial use with conditionsDetails
Video2025

Ovi

Character.AI · USA

Generates video together with sound and speech from text or an image: two branches (video based on Wan 2.2 and a 5B audio branch) run in sync. Needs 24–32 GB of GPU memory.

  • Short clips with talking characters
  • Animating an image with voice-over
  • Ad scene prototypes
Sizes
11B
Hardware
from: 1 GPU
Commercial use allowedDetails

Text 12

TextGGUF2025–2026

Kimi

Moonshot AI · China

Very large Moonshot MoE models for agentic work. K3 (2.8 trillion parameters) was the largest open model at release, with up to 1M tokens of context and image understanding; K2.7-Code is built for programming. On November 4 the weights of K3.1 Pro were opened (2.8 trillion parameters, 104B active, 1M tokens of context, images): a refinement of K3 for programming, 3D and financial research; by the tests of Moonshot itself, Terminal-Bench 4.0 is 65.7%. The weights are in MXFP4, about 1.56 TB, run through vLLM and SGLang, and need a serious cluster.

  • Multi-step agents: search, data collection, reports
  • In-depth document analysis
  • Help for developers
Sizes
16B-A3B – 2.8T-A104B
Hardware
from: 1 GPU
Commercial use with conditionsDetails
Text2026

Nanbeige 4.3

Nanbeige · China

A 33B model that the authors call the senior teacher for future lightweight agents: they plan to train small models (for example 3B) on it. It is strong in programming and office tasks: on the JobBench test across 35 professions it beats the baseline in most, according to the authors. Context up to 512 thousand tokens, reasoning can be switched on and off, and tool calling works best in the XML format. The authors compare it with Qwen3.8-27B, Muse-Glimmer-30B and Gemma 4 31B and say it leads on most tests; the figures are their own. A base version is published too.

  • an agent for programming and working with tools
  • office tasks by profession: reports, tables, document analysis
  • a base model for your own fine-tuning
Sizes
33B
Hardware
from: 1 GPU
Commercial use allowedDetails
Text2026

Raon-1 mini

KRAFTON AI · South Korea

A model from the Korean KRAFTON (the publisher of PUBG) with 30B parameters, 4B active, trained from scratch. It is strong in Korean conversation and agentic tasks: it can reason and call tools, and works with Hermes Agent, Pi and OpenCode. Training languages: Korean, English, Japanese, Chinese, Spanish, French, Thai, and Russian is not among them. There are compressed MXFP8 and NVFP4 versions and an auxiliary model for faster generation. This is a trial release (preview), and the specifications may change before the final one. The weights are issued on request on Hugging Face, and the benchmark figures are the authors' own.

  • a Korean-language assistant and customer support
  • an agent with tool calling on your own server
  • game and media projects for a Korean audience
Sizes
30B-A4B
Hardware
from: 1 GPU
Commercial use allowedDetails
TextGGUF2025–2026

Step

StepFun · China

StepFun MoE models built for fast, low-cost work: with 196 billion parameters, Step-3.5/3.7-Flash use about 11 billion per token. Compact Step3-VL-10B for images and voice Step-Audio 2 mini are available. The flagship Step 5 Preview (600B, 27B active, up to 1M-token context, text and images as input) is built for agents, programming and finance tasks.

  • High-load agents
  • Analysis of documents with diagrams and screenshots
  • Help for developers
Sizes
8B to 600B
Hardware
from: 1 GPU
Commercial use with conditionsDetails
TextGGUF2025–2026

LLaDA

Renmin University of China (GSAI) and Ant Group (inclusionAI) · China

Diffusion language models: text is written in blocks and then refined rather than word by word, which speeds up generation. LLaDA2.2 can edit what it has written and targets agents. LLaDA-Image is a separate product.

  • Fast generation of code and text
  • Agent scenarios with long context
  • Research into alternatives to standard LLMs
Sizes
8B – 100B (MoE)
Hardware
from: 1 GPU
Commercial use allowedDetails
TextOllama2026

Muse Glimmer

Meta Superintelligence Labs · USA

An open Meta model for agents on affordable hardware: distilled from the closed Muse Spark, understands text and images, trained on 100+ languages.

  • Agents with tool calling
  • Analysis of screenshots, charts and documents
  • Multilingual assistant
Sizes
30B
Hardware
from: 1 GPU
Commercial use allowedDetails
TextOllama2023–2026

Solar

Upstage · South Korea

Models from Korea's Upstage. Solar Open 2 is built for office document work: 250 billion parameters, 15 billion active; languages are English, Korean and Japanese.

  • Working with office documents
  • Agents for routine tasks
  • Help for developers
Sizes
10.7B – 250B-A15B
Hardware
from: 1 GPU
Commercial use with conditionsDetails
TextGGUF2024–2026

INTELLECT

Prime Intellect · USA

Models trained in a distributed way on GPUs from around the world. INTELLECT-3 (106B) is further trained with reinforcement learning for math, code and agents.

  • Reasoning and math tasks
  • Programming help
  • Agents with tool calling
Sizes
10B – 106B-A12B
Hardware
from: 1 GPU
Commercial use allowedDetails
TextOllama2025

gpt-oss

OpenAI · USA

OpenAI's first open models since GPT-2. Reasoning and tool calling; the smaller version fits on a single GPU.

  • AI agent that calls internal systems
  • Answers based on internal policies
  • Drafts of emails and reports
Sizes
20B, 120B
Hardware
from: 1 GPU
Commercial use allowedDetails
TextGGUF2025

Seed-OSS

ByteDance · China

An open ByteDance 36B model with up to 512K tokens of context and an adjustable thinking budget. Fits on a single powerful GPU.

  • Analysis of long documents
  • Agents with tools
  • Corporate assistant
Sizes
36B
Hardware
from: 1 GPU
Commercial use allowedDetails
TextOllama2024

Athene

Nexusflow · USA

Fine-tuned Llama 3 and Qwen 2.5 models from Nexusflow. Athene-V2-Agent is specially trained for function calling and agent scenarios. Commercial use is prohibited.

  • Research on agents and function calling
  • Comparison with commercial models
  • Experiments with a chat assistant
Sizes
70B – 72B
Hardware
from: 1 GPU
Non-commercial onlyDetails
TextRUGGUFNot maintained2023

ruGPT-3.5

Sber (ai-forever) · Russia

Sber's 13-billion-parameter base Russian model; GigaChat grew out of its fine-tuned version. Continues texts in Russian and English, context only 2048 tokens; today useful as a base for narrow fine-tuning.

  • Base for fine-tuning on a narrow Russian-language task
  • Generating template Russian texts
  • Experiments with Russian-language models without license restrictions
Sizes
13B
Hardware
from: 1 GPU
Commercial use allowedDetails

3D 12

3D2026

Idol

Sperid Labs · not stated

From a single picture of an object it makes a three-dimensional model in the form of Gaussian splats (the 3DGS .ply format). It takes a render, a generated image or a photo, removes the background itself, frames the object and fills in sides that the shot does not show, up to a million Gaussians. It runs on an ordinary graphics card with 7.2 GB of memory, in two seconds according to the authors. There are modes for renders and for photographs. The result opens in any splat viewer. The weights are Apache 2.0, but running it needs two gated encoders (DINOv3 and the VAE from FLUX.2-dev) with their own terms, which affects commercial use.

  • quick 3D previews of products from one photo for cards and catalogs
  • three-dimensional drafts from generated images
  • object prototypes for games and AR
Sizes
about 3 GB of weights
Hardware
from: 1 GPU
Commercial use with conditionsDetails
3D2026

Kimodo

NVIDIA · USA

A diffusion model that creates three-dimensional body animation from a text description and set conditions: poses, positions of hands and feet, a route and points to pass through. There are versions for the SOMA skeleton (30 joints), the Unitree G1 robot (34 joints) and the SMPL-X human (22 joints). Version 1.2 is trained on a corrected set: in 1.1 some test recordings leaked into training, so the new version is the one to compare against other models. According to the authors, foot skating is reduced and text following is slightly weaker. The data of some variants is proprietary (Bones Rigplay), while the open variants are trained on Bones-SEED.

  • animating characters for games and video from a description
  • motions for humanoid robots (the Unitree G1 skeleton)
  • motion drafts with a given route and poses
Sizes
not stated
Hardware
from: 1 GPU
Commercial use allowedDetails
3D2025–2026

VGGT

Meta and the University of Oxford (VGG) · USA / UK

Reconstructs a 3D scene from one, several or hundreds of photos in seconds: camera positions, depth and a point cloud. Best Paper at CVPR 2025.

  • 3D model of a room or object from a photo series
  • Camera pose estimation for photogrammetry
  • Point cloud for measurements and comparison with the plan
Sizes
about 1.2B
Hardware
from: 1 GPU
Commercial use with conditionsDetails
3D2025–2026

HY-World (HunyuanWorld)

Tencent · China

Generates whole 3D worlds and scenes from text or an image that you can walk through. The second version builds a scene from video and photos.

  • 3D scenes for games and virtual tours
  • Backgrounds and environments for video production
  • Draft locations for simulations
Sizes
set of several models
Hardware
from: 1 GPU
Commercial use with conditionsDetails
3D2025

HY-Motion

Tencent Hunyuan · China

Generates 3D human motion animation from a text description: the skeletal animation is ready for 3D editors and game engines. Understands English and Chinese.

  • Character animation from a text description
  • Draft animation for games and videos
  • Motion library for avatars
Sizes
0.46B – 1B
Hardware
from: 1 GPU
Commercial use with conditionsDetails
3D2025

MapAnything

Meta and Carnegie Mellon University · USA

A single model builds a metric 3D reconstruction from photos, and uses camera, depth or pose data when available. One weights variant is under Apache 2.0.

  • 3D reconstruction of an object or room from photos
  • Exporting the scene to COLMAP format for further processing
  • Depth and camera pose estimation
Sizes
about 1.2B
Hardware
from: 1 GPU
Commercial use allowedDetails
3D2025

Pi3 (π³)

Shanghai AI Lab · China

Reconstructs a 3D scene and camera positions from a set of photos or a video without relying on a "reference" frame. Pi3X gives smoother point clouds and approximate scale in meters.

  • 3D scene reconstruction from video
  • Camera pose estimation from frames
  • Point clouds for research and prototypes
Sizes
0.96B – 1.4B
Hardware
from: 1 GPU
Non-commercial onlyDetails
3DGGUF2024–2025

TRELLIS

Microsoft · USA

One of the strongest open 3D models: from an image or text it produces a textured mesh or a Gaussian scene. TRELLIS.2 is noticeably more detailed than the first version.

  • 3D models of products and interiors from photos
  • Assets for games and AR/VR
  • Prototypes for 3D printing
Sizes
up to 4B (TRELLIS.2)
Hardware
from: 1 GPU
Commercial use allowedDetails
3DGGUF2025

SAM 3D

Meta · USA

Reconstructs the 3D shape of an object or a human body from one ordinary photo, even when the object is partly hidden. Two models: Objects and Body.

  • 3D model of an item from a catalog photo
  • Estimating body pose and shape from a photo
  • Try-on and AR scenarios
Sizes
size not stated on the model card
Hardware
from: 1 GPU
Commercial use with conditionsDetails
3D2024–2025

Hunyuan3D

Tencent · China

Tencent's open 3D line: shape and texture from an image, at the level of paid services. Omni adds control of pose and shape, Part splits a model into parts.

  • Textured 3D product models
  • Characters and objects for games
  • Splitting a model into parts for printing
Sizes
set of models: shape and textures
Hardware
from: 1 GPU
Commercial use with conditionsDetails
3D2024–2025

Stable Fast 3D / SPAR3D

Stability AI · UK

Stability AI models that turn a single photo into a textured 3D model in about a second. SPAR3D lets you adjust the shape through a point cloud.

  • 3D product cards from photos
  • Assets for games and AR
  • Quick mock-ups for design
Sizes
1B – 2B
Hardware
from: 1 GPU
Commercial use with conditionsDetails
3DNot maintained2024

InstantMesh

Tencent ARC · China

Builds a 3D mesh from a single image in about 10 seconds: first it draws the object from several angles, then assembles the model from them.

  • 3D model of an object from a photo
  • Assets for games and visualizations
  • Prototypes for 3D printing
Sizes
size not stated on the model card
Hardware
from: 1 GPU
Commercial use allowedDetails

Virtual try-on 9

Virtual try-on2026

FASHN VTON

FASHN AI · Israel

A rare open try-on model with a commercial license: mask-free, accepts a photo of the item on a model or a flat lay. Weights are about 2 GB.

  • Product cards on a model without a photo shoot
  • Fitting room on a store website
  • Catalog from flat-lay clothing photos
Sizes
972M
Hardware
from: 1 GPU
Commercial use allowedDetails
Virtual try-on2025

OmniTry

Kunbyte AI · China

Try-on beyond clothing: glasses, earrings, bags, hats, watches and other accessories. Works without a mask. Needs a GPU with 28 GB or more.

  • Trying accessories and jewelry on a photo
  • Product cards with an accessory on a model
  • Online fitting room for eyewear and jewelry
Sizes
add-on for FLUX.1 Fill dev 12B
Hardware
from: 1 GPU
Commercial use with conditionsDetails
Virtual try-on2025

FastFit

LavieAI and Sun Yat-sen University · China

Tries on several items at once: top, bottom, shoes, bag. Faster than earlier models thanks to caching. Non-commercial license.

  • Building an outfit from several products on one model
  • "Build a look" pilot on a website
  • Lookbook prototypes without a shoot
Sizes
based on SD 1.5 inpainting
Hardware
from: 1 GPU
Non-commercial onlyDetails
Virtual try-on2025

Any2AnyTryon

Beijing University of Posts and Telecommunications and others · China

An all-round FLUX-based apparel toolkit: try-on, generating a model wearing a given item, and "taking off" an item from a person into a separate product photo.

  • Try-on from a product photo
  • Photo of a model wearing an item from a text description
  • Clean product photo extracted from a shot of a person
Sizes
add-ons (LoRA) for FLUX.1 dev 12B
Hardware
from: 1 GPU
Non-commercial onlyDetails
Virtual try-on2024

FitDiT

Tencent and Fudan University · China

Tencent's transformer-based try-on: more accurately reproduces fabric texture, fine prints and garment length. Non-commercial license.

  • Try-on of items with complex prints and textures
  • Test product cards on a model
  • Checking length and fit on a photo
Sizes
based on SD3
Hardware
from: 1 GPU
Non-commercial onlyDetails
Virtual try-on2024

Leffa

Meta AI (with King's College London) · USA

Meta's model for virtual try-on and changing a person's pose in a photo. Carefully transfers fine fabric details and lettering. MIT license, but the training data is non-commercial.

  • Trying clothes on a model photo
  • Changing the model pose in an existing shot
  • Adding extra angles for a product card
Sizes
based on Stable Diffusion
Hardware
from: 1 GPU
Commercial use with conditionsDetails
Virtual try-onNot maintained2024

IDM-VTON

KAIST and OMNIOUS.AI · South Korea

One of the best-known open virtual try-on models: moves a garment from a product photo onto a photo of a person, keeping prints and logos well. Non-commercial license.

  • Pilot of a fitting room on a store website
  • Prototype product cards on a model without a photo shoot
  • Comparison with commercial try-on services
Sizes
based on SDXL
Hardware
from: 1 GPU
Non-commercial onlyDetails
Virtual try-onNot maintained2024

OOTDiffusion

Xiao-i Research · China

An early popular open try-on model: one version for upper-body garments, another for full-length outfits. Non-commercial license.

  • Trying tops on a model photo
  • Full-length try-on: tops, bottoms, dresses
  • Fitting room prototype for testing
Sizes
based on Stable Diffusion
Hardware
from: 1 GPU
Non-commercial onlyDetails
Virtual try-onNot maintained2024

StableVITON

KAIST · South Korea

A research try-on model from CVPR 2024, one of the first built on Stable Diffusion. Now mostly used as a comparison baseline.

  • Pilot of upper-body garment try-on
  • Comparing quality of different try-on models
  • Training your own try-on on the open code
Sizes
based on Stable Diffusion
Hardware
from: 1 GPU
Non-commercial onlyDetails

Robotics 8

Robotics2025–2026

NVIDIA Cosmos

NVIDIA · USA

"World" models for robots and self-driving vehicles: they generate realistic video of physical scenes and predict actions. Cosmos 3 combines understanding, generation and control. In October the surgical versions came out: Cosmos3-H-Embodied-Nano (16B, one model for predicting video, actions and policy on surgical robots) and Cosmos3-H-Dreams (a real-time world for simulating operations with the Versius robot). They are under the NVIDIA Open Model License.

  • Synthetic video for training robots and self-driving vehicles
  • Testing scenarios in simulation
  • Robot control (Policy versions)
Sizes
2B – 65B
Hardware
from: 1 GPU
Commercial use allowedDetails
Robotics2025–2026

GigaBrain

GigaAI · China

A robot control model trained mostly on synthetic data from a world model. It reduces spending on collecting data from real robots.

  • Controlling a robot arm
  • Fine-tuning on a small amount of your own data
  • Sorting and assembly pilots
Sizes
3.5B
Hardware
from: 1 GPU
Commercial use allowedDetails
Robotics2026

LingBot-VLA

Ant Group (Robbyant) · China

A robot control model from Ant Group trained on a large volume of data from real robots. Version 2.0 works with different types of robot arms.

  • Controlling a two-armed robot
  • Fine-tuning for your own operation
  • Assembly and sorting pilots
Sizes
4B – 6B
Hardware
from: 1 GPU
Commercial use allowedDetails
Robotics2026

Xiaomi Robotics

Xiaomi · China

Open robot control models from Xiaomi. Robotics-1 is designed for household and kitchen tasks, U0 combines scene understanding and action.

  • Controlling a robot arm by command
  • Household and service scenarios
  • Base for fine-tuning
Sizes
4B – 5B
Hardware
from: 1 GPU
Commercial use allowedDetails
Robotics2025–2026

MolmoAct

Ai2 (Allen Institute for AI) · USA

A fully open robot control model that first "reasons" about space and trajectory, then acts. Its reasoning can be checked.

  • Controlling a robot arm with explainable steps
  • Fine-tuning for your own robot
  • Research pilots
Sizes
5B – 8B
Hardware
from: 1 GPU
Commercial use allowedDetails
Robotics2025–2026

NVIDIA Isaac GR00T

NVIDIA · USA

NVIDIA's foundation model for humanoid robots and robot arms: it sees, understands a command and outputs movements. Built into the Isaac ecosystem.

  • Controlling humanoid robots
  • Fine-tuning for your own robotic cell
  • Training in simulation with transfer to a real robot
Sizes
2B – 3B
Hardware
from: 1 GPU
Commercial use with conditionsDetails
Robotics2025

π0 / π0.5 (openpi)

Physical Intelligence · USA

Robot control models from Physical Intelligence: folding laundry, tidying up, handling objects. π0.5 copes better in unfamiliar settings.

  • Controlling robot arms and two-armed robots
  • Fine-tuning for your own operations
  • Pilots for automating manual work
Sizes
about 3B
Hardware
from: 1 GPU
Commercial use with conditionsDetails
Robotics2024–2025

OpenVLA

Stanford, Berkeley and partners · USA

The first large open vision-language-action model: a robot arm carries out commands like "put the apple in the bowl". OFT makes it several times faster.

  • Controlling a robot arm by text command
  • Pilots for robotizing simple operations
  • Base for fine-tuning to your own robot
Sizes
7B
Hardware
from: 1 GPU
Commercial use allowedDetails

Avatars 7

AvatarsGGUF2025–2026

LongCat-Video-Avatar

Meituan · China

Audio-driven talking people built on LongCat-Video. Version 1.5 is production-ready: stable long videos in Chinese and English.

  • News or course presenter videos
  • Promo videos with a talking character
  • Singing and voice-over
Sizes
based on LongCat-Video 13.6B
Hardware
from: 1 GPU
Commercial use allowedDetails
Avatars2024–2026

Hallo

Fudan University · China

A series of audio-driven talking portraits: from short clips to hour-long 4K videos. Hallo-Live is built for real-time use.

  • Presenter video from a photo and audio
  • Long training videos
  • Live avatar
Sizes
about 1B – 5B
Hardware
from: 1 GPU
Commercial use allowedDetails
Avatars2026

daVinci-MagiHuman

SII-GAIR and Sand.ai · China

Generates video of a talking person with sound in one go: a single transformer processes text, video and audio. Speech in 7 languages; Russian is not among them. Fast distilled versions are available.

  • Presenter video from a script
  • Ad videos with a talking character
  • Training videos with a narrator
Sizes
15B
Hardware
from: 1 GPU
Commercial use allowedDetails
Avatars2024–2026

EchoMimic

Ant Group · China

Ant Group's talking avatars: the face and, from V2, hand gestures. V3-Flash produces video in 8 steps and fits into 12 GB of GPU memory.

  • Presenter video from a photo and voice
  • Avatar with gestures for presentations
  • Voiced characters
Sizes
up to 1.3B
Hardware
from: 1 GPU
Commercial use allowedDetails
AvatarsGGUF2025–2026

Live Avatar

Alibaba (Quark) · China

A real-time streaming avatar of unlimited length. Suits live broadcasts and dialogue, but needs powerful server hardware.

  • Live avatar for customer dialogue
  • Endless broadcasts with a presenter
  • Interactive characters
Sizes
14B
Hardware
from: 1 GPU
Commercial use allowedDetails
Avatars2025

MultiTalk / InfiniteTalk

Meituan · China

Dubbing and talking characters built on Wan: MultiTalk handles dialogue between several people, InfiniteTalk re-dubs videos of any length with facial and body motion.

  • Video dubbing with matched facial expressions
  • Dialogue between two characters from audio
  • Long videos with a presenter
Sizes
14B
Hardware
from: 1 GPU
Commercial use allowedDetails
AvatarsGGUF2025

HunyuanVideo-Avatar

Tencent · China

Talking characters built on HunyuanVideo: conveys emotions from the voice, handles several characters and different styles.

  • Presenter video from a photo and audio
  • Scenes with several speakers
  • Cartoon characters
Sizes
about 13B
Hardware
from: 1 GPU
Commercial use with conditionsDetails

Image + text 7

Image + text2026

Interfaze 1 Lite

Interfaze · not stated

One model for developers: reads documents and tables, transcribes speech with speaker labels, finds objects in an image and answers strictly by a JSON schema you supply. Inside is a reasoning core and a set of narrow specialist models.

  • parsing documents and tables with text coordinates
  • transcribing recordings with speaker labels
  • extracting data from documents into a given JSON schema
Sizes
a model that needs 80 GB of video memory (the core is based on Qwen3.8-27B)
Hardware
from: 1 GPU
Commercial use with conditionsDetails
Image + text2026

LensVLM

Apple · USA

Reads long documents differently: it turns pages into downscaled images and expands the full text only for the pages needed to answer. A long document then costs far less context.

  • Answering questions over multi-page documents
  • Finding the right page in a large report or case file
  • Working with scans and PDFs without a separate OCR step
Sizes
9B
Hardware
from: 1 GPU
Non-commercial onlyDetails
Image + text2026

MOSS-VL

OpenMOSS (Fudan University) · China

An image + video + text model focused on long videos and precise linking of events to timestamps. A Realtime version handles live video streams.

  • Analyzing long videos and finding events by time
  • Real-time streaming video analysis
  • Understanding photos and documents
Sizes
about 11B
Hardware
from: 1 GPU
Commercial use allowedDetails
Image + text2026

Qwen3-VL Resume Parser

Sukhrob Nurali · not disclosed

A fine-tuned Qwen3-VL-8B reads resume pages as images and returns a 23-field JSON record. The author states plainly that the model is not meant for automated decisions about candidates; a human decides.

  • Moving a resume from PDF into a candidate record
  • Filling a candidate database without manual typing
  • Parsing resumes with different layouts and styling
Sizes
8B, a fine-tune of Qwen3-VL-8B-Instruct
Hardware
from: 1 GPU
Commercial use allowedDetails
Image + textRU2025

A-Vision (Авито)

Avito Tech · Russia

Avito's Russian-language model that understands images: describes photos, answers questions about an image, reads text on it. Based on Qwen2.5-VL, faster in Russian than the original.

  • Product descriptions from photos in Russian
  • Checking that a photo matches its description
  • Reading brands and text in images
Sizes
7.4B
Hardware
from: 1 GPU
Commercial use allowedDetails
Image + textGGUF2025

Kimi-VL

Moonshot AI · China

An efficient MoE vision model (16B, 3B active) with a long context and a reasoning version. Handles long documents and video well.

  • Analysing long PDFs and presentations
  • Answering questions about video
  • Operating interfaces from screenshots
Sizes
16B-A3B
Hardware
from: 1 GPU
Commercial use allowedDetails
Image + textNot maintained2023–2024

IDEFICS

Hugging Face · France / USA

Open vision models from Hugging Face that reproduced the closed Flamingo. Idefics3 became the basis for the compact SmolVLM line.

  • Answering questions about images
  • Analysing documents and screenshots
  • A base for fine-tuning
Sizes
8B – 80B
Hardware
from: 1 GPU
Commercial use allowedDetails

Computer-use agents 6

Computer-use agents2026

UI-Mate

Tencent · China

An open agent for operating a desktop: looks at live screenshots, reasons and outputs clicks and keyboard input. You can show a workflow once and the agent carries it over to a new task. An external environment is needed to execute the actions.

  • automating work in programs with no API
  • repeating a workflow from one shown example
  • testing interfaces
Sizes
27B
Hardware
from: 1 GPU
Commercial use allowedDetails
Computer-use agents2026

AREX-2

Beijing Academy of Artificial Intelligence (BAAI) · China

An agent model of 27B (based on Qwen3.8-27B) for long tasks: it proposes a solution, measures the result, analyzes errors and revises the solution over several rounds. A 262K-token context. Stated as a research model.

  • iterative improvement of code and algorithms
  • machine learning tasks with a verifiable result
  • deep search and gathering information on a topic
Sizes
27B
Hardware
from: 1 GPU
Commercial use allowedDetails
Computer-use agents2025–2026

OpenCUA / Qwen-CUA

XLANG Lab (University of Hong Kong) · China

Fully open desktop agents: weights, data and training code. They work on Windows, macOS and Linux; the latest Qwen-CUA controls a computer with ordinary clicks and keystrokes.

  • Working in desktop software without an API
  • Moving data between systems
  • Running user scenarios for tests
Sizes
7B – about 400B (MoE)
Hardware
from: 1 GPU
Commercial use allowedDetails
Computer-use agentsGGUF2026

EvoCUA

Meituan · China

Meituan's computer-control agent, trained on a large number of simulated tasks in desktop software. It outputs clicks and keyboard input.

  • Working in office and legacy software without an API
  • Moving data between systems
  • Running test scenarios
Sizes
8B – 32B
Hardware
from: 1 GPU
Commercial use allowedDetails
Computer-use agents2023–2025

CogAgent / AutoGLM

Zhipu AI (Z.ai) and Tsinghua University · China

One of the first open models for controlling an interface from a screenshot; its successor, AutoGLM-Phone, works in Android smartphone apps.

  • Automating actions in mobile apps
  • Working in web interfaces without an API
  • Testing apps against scenarios
Sizes
9B – 18B
Hardware
from: 1 GPU
Commercial use with conditionsDetails
Computer-use agentsGGUF2025

Magma

Microsoft Research · USA

An agent model that plans actions both in an interface (buttons on screen) and for a robot (arm movements). For now more of a research base than a finished product.

  • Pilots in interface control
  • Research projects spanning screens and robotics
  • Analyzing screenshots with an action plan
Sizes
8B
Hardware
from: 1 GPU
Commercial use allowedDetails

Visual document search 6

Visual document search2025

Nomic Embed Multimodal / ColNomic

Nomic AI · USA

Search across PDF pages and scans as images. The cards list English, Italian, French, German and Spanish — Russian is not among them.

  • Search across an archive of scans and PDFs
  • Search across tables and diagrams inside documents
  • Picking pages for an AI assistant answer
Sizes
3B and 7B
Hardware
from: 1 GPU
Commercial use with conditionsDetails
Visual document search2025

LlamaIndex vdr

LlamaIndex · USA

A small model for searching document pages as images, from the team behind a popular RAG framework. The card lists English, Italian, French, German and Spanish.

  • Search across scans and PDFs without OCR
  • Search across invoices, acts and contracts
  • Picking pages for an AI assistant answer
Sizes
2B (based on Qwen2-VL)
Hardware
from: 1 GPU
Commercial use allowedDetails
Visual document search2024

GME (General Multimodal Embedding)

Alibaba (Tongyi Lab) · China

One vector for text, for an image and for a text-image pair: a single model can find a product by photo, a document page by question and an image by description. The card lists English and Chinese.

  • Finding a product by photo
  • Search across a catalogue of images and cards
  • Search across document pages as images
Sizes
2B and 7B
Hardware
from: 1 GPU
Commercial use allowedDetails
Visual document search2024

MonoQwen2-VL (LightOn)

LightOn · France

A reranker for document pages as images: after a visual search it reorders the found pages by how well they answer the question. The card does not state the languages.

  • Refining search results over scans and PDFs
  • Selecting pages before an AI assistant answers
  • Sorting retrieved slides and reports
Sizes
2B (based on Qwen2-VL)
Hardware
from: 1 GPU
Commercial use allowedDetails
Visual document search2024

VLM2Vec

TIGER-Lab · Canada

Turns an image-plus-text model into an embedding model: one vector for a page, a diagram or a captioned photo. The card states English.

  • Search across a mixed archive of texts and images
  • Search across document pages as images
  • Finding similar cards and illustrations
Sizes
about 4B (based on Phi-3.5-V)
Hardware
from: 1 GPU
Commercial use allowedDetails
Visual document search2024

DSE (Document Screenshot Embedding)

University of Waterloo, Tevatron project · Canada

Searches page screenshots: the page is not OCRed but turned into a single vector, so the index is more compact than with late-interaction models. The card lists English and French.

  • Search across scans and PDFs without OCR
  • Search across presentations and reports with complex layouts
  • Picking pages for an AI assistant answer
Sizes
2B (based on Qwen2-VL)
Hardware
from: 1 GPU
Commercial use allowedDetails

Code 5

Code2026

Flint

Recursive · not stated

An open expert model (39.8B parameters, 2.5B active per token) for coding agents: it digs into someone else's repository, fixes bugs, writes functions and tests, and can reason and call tools over several steps. Trained from scratch, 262 thousand token context, text only. It weighs about 80 GB, and the authors run it on a single B300 or H200 card through SGLang. By the authors' own tests: Terminal-Bench 2.1 54.9%, SWE-bench Pro v2 89.7%, SWE-bench Verified 74.2%. Part of the comparisons come from other developers' own reports, and the test conditions differ between sources.

  • an agent for fixing bugs and extending a repository in the terminal
  • writing tests and new functions from a description
  • a local developer assistant on your own server
Sizes
39.8B-A2.5B
Hardware
from: 1 GPU
Commercial use allowedDetails
CodeOllama2026

Ornith

DeepReinforce · not disclosed

Models for agentic development: they build their own plan and scaffolding for a task and execute it in the terminal. Fine-tuned from Qwen 3.5 and Gemma 4; work with Claude Code, OpenHands and similar tools. In October Ornith-1.5-122B (MoE, MIT) came out: according to the authors it is stronger and faster than the previous ones, writes shorter and is meant for a single DGX-class card; compressed FP8, NVFP4, GGUF and MLX versions exist. The figures come from the authors.

  • A developer agent in the terminal
  • Fixing bugs from a task description
  • Understanding and extending a large repository
Sizes
9B – 397B
Hardware
from: 1 GPU
Commercial use allowedDetails
Code2025–2026

KAT-Coder / KAT-Dev

Kwaipilot (Kuaishou) · China

Kuaishou models for agentic development, trained to solve real tasks in repositories. KAT-Coder-V2.5-Dev (35B, 3B active) is the open version of their closed flagship.

  • An agent that fixes tasks in the repository
  • Code generation and refactoring
  • Automating routine development tasks
Sizes
32B – 72B, 35B-A3B
Hardware
from: 1 GPU
Commercial use allowedDetails
CodeGGUF2026

Laguna

Poolside · USA

Models for agentic programming: they edit code in a repository on their own. The small XS runs on a Mac with 36 GB of memory; S 2.1 has a 1M-token context.

  • Coding agent for in-house development
  • Bug fixing and code improvements
  • Working with large codebases
Sizes
33B-A3B – 225B-A23B
Hardware
from: 1 GPU
Commercial use allowedDetails
CodeRUOllama2025

Devstral

Mistral AI (with All Hands AI) · France

Mistral models for agentic development: they read the repository, edit files and run commands on their own. The 24B version fits on a single GPU.

  • A developer agent that fixes tickets from the tracker
  • Extending internal systems from a description
  • Automating routine code edits
Sizes
24B – 123B
Hardware
from: 1 GPU
Commercial use with conditionsDetails

Medicine 5

Medicine2026

Carbon-X

Hugging Face (HuggingFaceBio) · USA

An open 14.8B language model for working with DNA, built on Qwen3-14B. You send a raw sequence right in an ordinary chat, and it answers in English: it judges whether a mutation is benign or pathogenic, finds regions that control gene activity, assigns a sequence to an organism group, predicts the next bases and explains its reasoning. DNA and text go in one stream, with DNA encoded in groups of six bases. By the authors' tests, on five sets it scores 66.7% without reasoning, 27 points above Qwen3-14B, and costs about 1,800 times less than GPT-6.1 Sol at a close result. It fits on a single 80 GB card, so sensitive data can stay with you. Viral data was excluded from training because of the risks. The conclusions are computed and do not replace laboratory verification.

  • assessing DNA variants: benign or pathogenic
  • finding regions that control gene activity
  • a research assistant for biologists on your own server
Sizes
14.8B
Hardware
from: 1 GPU
Commercial use allowedDetails
Medicine2026

NV-Reason-CT

NVIDIA · USA

A language model for computed tomography: looks at a whole chest or abdominal CT volume, not separate slices, writes structured reports and answers questions with reasoning. English only.

  • a draft CT report for a physician
  • yes-or-no answers about a scan with reasoning
  • research in medical imaging
Sizes
4B (based on Qwen3.5-4B with a 3D encoder)
Hardware
from: 1 GPU
Commercial use allowedDetails
MedicineGGUF2025–2026

AntAngelMed

Ant Healthcare (Ant Group) and Zhejiang Provincial Medical Information Center · China

A large medical MoE model based on Ling-flash-2.0: 100B parameters with 6B active, so it answers quickly. Does not replace a doctor; decisions are made by a specialist.

  • Reference answers to staff on clinical questions
  • Draft discharge summaries for a doctor to review
  • Searching medical literature
Sizes
100B-A6B
Hardware
from: 1 GPU
Commercial use allowedDetails
MedicineNot maintained2023

Clinical Camel

Bo Wang's lab (University of Toronto, Vector Institute) · Canada

An early medical model on Llama 2 70B, trained on dialogues based on medical texts. Now mainly of research interest. Does not replace a doctor; decisions are made by a specialist.

  • Research pilots on medical dialogue
  • Training materials for staff
  • Comparison with newer medical models
Sizes
70B
Hardware
from: 1 GPU
Non-commercial onlyDetails
MedicineNot maintained2023

RadFM

Shanghai Jiao Tong University and Shanghai AI Lab · China

An early general-purpose radiology model: understands 2D and 3D images (CT, MRI) together with text. More of a research base. Does not replace a doctor; decisions are made by a specialist.

  • Research pilots on CT and MRI analysis
  • Hints for doctors when reviewing images
  • A base for fine-tuning on the clinic's own images
Sizes
size not stated on the model card
Hardware
from: 1 GPU
Commercial use with conditionsDetails

Weather and climate 5

Weather and climate2024–2026

Ai2 ACE2 (климатический эмулятор)

Allen Institute for AI (Ai2) · USA

A fast climate model emulator: simulates the atmosphere years and decades ahead on a single GPU. Coupled with an ocean model (SamudrACE) for long-term scenarios. The new HiRO-ACEv2 adds a diffusion model that sharpens the coarse 100 km grid to 3 km: precipitation, wind and pressure every 6 hours for decades ahead, worldwide or for a chosen region. By the estimate of Ai2, global 3 km data come out about 300 times cheaper than with the X-SHiELD model. For now it reproduces only the X-SHiELD climate, not real weather from the archive.

  • Decades-long climate scenarios to assess long-term asset risks
  • Large-scale what-if runs on temperature and precipitation
  • Preparing data for crop yield and energy demand models
Sizes
checkpoint of about 1.8 GB
Hardware
from: 1 GPU
Commercial use allowedDetails
Weather and climate2023–2026

Google DeepMind GraphCast / GenCast / WeatherNext 2

Google DeepMind · UK

Google DeepMind's family of global weather models: GraphCast (10-day forecast), GenCast (probabilistic ensemble) and WeatherNext 2 with cyclone forecasting. Since August 2026 the weights are cleared for commercial use.

  • Medium-range weather forecasts for planning shifts, voyages and deliveries
  • Probabilistic assessment of extreme weather for insurance portfolios
  • Tropical cyclone track forecasts for marine and port operations
Sizes
from lightweight 1° versions to full 0.25°
Hardware
from: 1 GPU
Commercial use allowedDetails
Weather and climate2024–2026

Microsoft Aurora

Microsoft Research · USA

A foundation model of Earth's atmosphere: global weather forecasts, plus separate versions for air quality and ocean waves. Computes a forecast in seconds instead of hours on a supercomputer.

  • Your own forecast of temperature, wind and precipitation for company locations
  • Sea state estimates for planning voyages and port operations
  • Air pollution forecasts for industrial sites
Sizes
about 1.3B (a small test version is available)
Hardware
from: 1 GPU
Commercial use allowedDetails
Weather and climate2022–2026

NVIDIA FourCastNet и модели Earth-2

NVIDIA · USA

NVIDIA's set of weather and climate models: global FourCastNet forecasts, downscaling to kilometers (CorrDiff), regional storm forecasts (StormCast), climate generation (cBottle, Atlas). Run via Earth2Studio.

  • Global forecasts followed by downscaling to the region you need
  • Short-term forecasts of thunderstorms and heavy rain for dispatch services
  • Generating many weather scenarios for stress tests
Sizes
98M – 2.5B (FourCastNet 3 — about 711M)
Hardware
from: 1 GPU
Commercial use with conditionsDetails
Weather and climate2024–2026

ECMWF AIFS

European Centre for Medium-Range Weather Forecasts (ECMWF) · Europe (intergovernmental organization)

ECMWF's weather neural network running operationally: a 15-day forecast four times a day, an ensemble version with 51 scenarios, and since version 2, ocean waves.

  • Running your own forecast from open initial data
  • Ensemble forecasts to estimate the probability of frost, downpours and storms
  • Wave forecasts for marine operations
Sizes
checkpoint of about 1 GB
Hardware
from: 1 GPU
Commercial use allowedDetails

Music and sound 4

Music and sound2026

HeartMuLa

HeartMuLa Team · not disclosed

An open model for generating songs with vocals in Chinese, English, Japanese, Korean and Spanish, plus a codec and a lyrics transcription model.

  • Songs and jingles from lyrics
  • Music for videos
  • Transcribing song lyrics
Sizes
3B
Hardware
from: 1 GPU
Commercial use allowedDetails
Music and soundGGUF2025–2026

YuE

M-A-P and HKUST · China

Generates full songs with vocals and accompaniment from lyrics and a style description: English, Chinese, Japanese, Korean.

  • Songs and jingles from lyrics
  • Demo versions of tracks
  • Music for videos
Sizes
0.5B – 7B
Hardware
from: 1 GPU
Commercial use with conditionsDetails
Music and sound2025–2026

ThinkSound / PrismAudio

Alibaba Tongyi (FunAudioLLM) · China

Generates and edits audio for video, text or audio, first "reasoning" about the scene with a multimodal model. PrismAudio is the next version for video-to-audio.

  • Audio for video based on the scene
  • Editing individual sounds in a track
  • Sound effects from a description
Sizes
size not stated on the model card
Hardware
from: 1 GPU
Commercial use allowedDetails
Music and sound2025

HunyuanVideo-Foley

Tencent Hunyuan · China

Generates studio-quality (48 kHz) audio for video from the picture and a text prompt: footsteps, impacts, ambience, in sync with the action on screen.

  • Foley and sound effects for video
  • Sound for AI-generated ads
  • Sound design for short videos
Sizes
not stated on the model card (weights about 10 GB; XL version with memory offloading)
Hardware
from: 1 GPU
Commercial use with conditionsDetails

Photo editing 4

Photo editing2025

SeedVR / SeedVR2

ByteDance Seed · China

ByteDance's video and photo restoration and upscaling. SeedVR2 does it in a single step, so it is noticeably faster than similar models. Commercial-friendly license.

  • Upscaling photos and video to 2K–4K
  • Restoring old videos and photos
  • Enhancing user photos before publishing
Sizes
3B – 7B
Hardware
from: 1 GPU
Commercial use allowedDetails
Photo editing2024

Flux.1-dev ControlNet Upscaler

Jasper AI · USA

A FLUX add-on for upscaling small and blurry images with detail reconstruction. Popular, but under the non-commercial FLUX dev license.

  • Upscaling small images with detail reconstruction
  • Enhancing generated images
  • Upscaling pilots for a catalog
Sizes
add-on for FLUX.1 dev 12B
Hardware
from: 1 GPU
Non-commercial onlyDetails
Photo editingNot maintained2024

SUPIR

XPixel Group (Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shanghai AI Lab and others) · China

Powerful SDXL-based restoration of badly damaged photos: it recreates details rather than just upscaling. Hardware-hungry; non-commercial license.

  • Restoring old and blurry photos
  • Upscaling with detail reconstruction
  • Archive restoration pilots
Sizes
based on SDXL, plus LLaVA 13B for captions
Hardware
from: 1 GPU
Non-commercial onlyDetails
Photo editingNot maintained2023

StableSR

S-Lab, Nanyang Technological University · Singapore

One of the first Stable Diffusion-based photo upscalers: restores realistic details. Non-commercial license.

  • Upscaling photos with detail reconstruction
  • Research restoration pilots
  • Comparison with classic upscalers
Sizes
based on SD 2.1
Hardware
from: 1 GPU
Non-commercial onlyDetails

Documents and OCR 4

Documents and OCR2026

Youtu-Parsing-Omni

Tencent · China

One compact model turns a document page, chart, flowchart, geometry figure, audio or video into structured JSON: with block coordinates, tables, formulas, speech transcription and timestamps.

  • parsing scans and PDFs with tables and formulas
  • turning charts and flowcharts into a table or diagram
  • transcribing audio and video with timestamps
Sizes
5B
Hardware
from: 1 GPU
Commercial use with conditionsDetails
Documents and OCR2026

jina-ocr-v1

Jina AI · Germany

Document parsing in a single model: a whole page becomes Markdown - text in correct reading order, tables and formulas in LaTeX. Built on DeepSeek-OCR, with only 0.6B of its 3.4B parameters active.

  • Converting scans and PDFs to Markdown
  • Recognizing tables and formulas
  • Parsing invoices, acts and reports
Sizes
3.4B-A0.6B
Hardware
from: 1 GPU
Non-commercial onlyDetails
Documents and OCRGGUF2025

olmOCR

Ai2 (Allen Institute for AI) · USA

A model and toolkit for converting PDFs into clean text at scale, preserving reading order, tables and formulas. Built to process millions of pages.

  • Bulk digitisation of a PDF archive
  • Converting contracts and reports into text
  • Preparing documents for search and RAG
Sizes
7B
Hardware
from: 1 GPU
Commercial use allowedDetails
Documents and OCRNot maintained2024

Kosmos-2.5

Microsoft · USA

Turns a scanned page into tagged text with block coordinates, or into markdown. Handy as the first step before parsing a resume. A human makes the decision about a candidate; automatic screening without review must not be used.

  • Converting resume scans into text that keeps its structure
  • Preparing documents for field extraction
  • Digitising paper forms
Sizes
about 1.4B
Hardware
from: 1 GPU
Commercial use allowedDetails

Biology and chemistry 4

Biology and chemistry2022–2026

OpenFold / OpenFold3

AlQuraishi Lab (Columbia University) and the OpenFold consortium · USA

A fully open reproduction of AlphaFold 2 and then AlphaFold 3 under Apache 2.0, with training data. OpenFold3 predicts complexes of proteins, nucleic acids and ligands.

  • Predicting structures of proteins and ligand complexes
  • Fine-tuning on the company's own data (training code is open)
  • An in-house structural analysis service without sending data outside
Sizes
a single set of weights per version
Hardware
from: 1 GPU
Commercial use allowedDetails
Biology and chemistry2024–2026

Arc Institute Evo / Evo 2

Arc Institute (with Together AI, Stanford, NVIDIA) · USA

DNA language models with context up to a million nucleotides: they assess the impact of mutations, annotate genomes and generate sequences. Evo 2 is trained on genomes from all domains of life.

  • Assessing the likely harmfulness of genetic variants for research
  • Annotating genomes of microorganisms and plants
  • Finding promising sequences in breeding and synthetic biology
Sizes
1B – 40B
Hardware
from: 1 GPU
Commercial use allowedDetails
Biology and chemistry2024–2025

Boltz (Boltz-1, Boltz-2, BoltzGen)

MIT (Jameel Clinic) and Recursion · USA

An open MIT-licensed alternative to AlphaFold 3: predicts structures of protein, DNA and small-molecule complexes; Boltz-2 estimates binding strength, BoltzGen designs new binding proteins.

  • Predicting how a candidate molecule binds to a target protein
  • Ranking compounds by predicted binding strength before synthesis
  • Designing binder proteins for a given target
Sizes
checkpoints of about 2 GB
Hardware
from: 1 GPU
Commercial use allowedDetails
Biology and chemistry2024

AlphaFold 3

Google DeepMind and Isomorphic Labs · UK

The reference model for the structure of biomolecules and their complexes. Weights are provided for non-commercial research only; companies need commercial access via Google Cloud or open alternatives (Boltz, OpenFold3).

  • Academic research on protein and complex structures
  • Benchmarking open alternatives against the reference on your own targets
Sizes
a single set of weights
Hardware
from: 1 GPU
Non-commercial onlyDetails

Finance 4

Finance2025

Fino1 / Fin-o1

The Fin AI · international project

Models that spell out their reasoning before answering financial questions with numbers and tables. Not investment advice: decisions are made by a specialist.

  • Calculation questions on statements with the working shown
  • Working through tasks with tables and numbers from documents
  • Checking calculations made by hand
Sizes
8B и 14B
Hardware
from: 1 GPU
Commercial use with conditionsDetails
Finance2023–2024

XuanYuan

Du Xiaoman (Duxiaoman-DI) · China

A large Chinese model family for the financial industry: advice, document reading and long texts up to 8k-16k. Not investment advice: decisions are made by a specialist.

  • Answering customer questions about banking products
  • Working through long financial documents
  • An internal assistant for a finance company regulations
Sizes
6B – 176B
Hardware
from: 1 GPU
Commercial use with conditionsDetails
Finance2023–2024

FinGPT

AI4Finance Foundation · USA

An open set of lightweight add-ons for ordinary language models that work with financial texts and news. Not investment advice: decisions are made by a specialist.

  • Assessing the tone of financial news and reports
  • Tagging mentions of companies and instruments in text
  • Preparing digests from a stream of business news
Sizes
adapters for 6B - 20B base models
Hardware
from: 1 GPU
Commercial use with conditionsDetails
FinanceNot maintained2023

DISC-FinLLM

Fudan-DISC, Fudan University · China

A financial assistant made of several fine-tuned experts: advice, calculations, document reading and knowledge-base search. Not investment advice: decisions are made by a specialist.

  • In-house advice on financial questions
  • Reading financial documents and news
  • Prompts for front-office staff
Sizes
13B
Hardware
from: 1 GPU
Commercial use with conditionsDetails

Drawings and CAD 4

Drawings and CADGGUF2025–2026

OmniSVG

Fudan University and StepFun · China

A family of models that produce SVG both from an image and from a text description, from a simple icon to a detailed illustration. For precise drawings the result has to be checked against the original.

  • Vectorising images and symbols into SVG
  • Creating icons and pictograms from a text description
  • Preparing vector graphics for web and print
Sizes
3B – 8B
Hardware
from: 1 GPU
Commercial use allowedDetails
Drawings and CAD2025

QZhou-Flowchart-VL

Kingsoft · China

It reads a flowchart from an image and first writes out its structure as JSON - blocks, links and branching conditions - and only then answers a question about the chart. The parse still has to be checked against the original.

  • Converting a flowchart image into a JSON structure
  • Answering questions about a process diagram
  • Digitising procedures and diagrams from presentations
Sizes
32B
Hardware
from: 1 GPU
Commercial use allowedDetails
Drawings and CAD2024–2025

Text2CAD

DFKI and RPTU Kaiserslautern · Germany

It builds a sequence of CAD modelling commands straight from a text description of a part, from a beginner phrase to a designer-level wording. The output is a draft; an engineer checks and signs off the document.

  • A draft part from a text description
  • Quickly trying out shape options at an early stage
  • Filling a library of standard parts
Sizes
size not stated on the model card
Hardware
from: 1 GPU
Non-commercial onlyDetails
Drawings and CAD2023–2024

PlankAssembly

Manycore Research · China

It assembles a 3D model of a product from three standard drawing projections - front, top and side - including cases where the lines in the drawing are incomplete. The output is a draft; an engineer checks and signs off the document.

  • Turning a three-view drawing into a 3D model
  • Digitising an archive of paper drawings for furniture and cabinets
  • Checking that the views agree with each other
Sizes
size not stated on the model card, several checkpoints published for different inputs
Hardware
from: 1 GPU
Commercial use with conditionsDetails

Text analysis 4

Text analysis2026

CLM (Contrastive Language Model)

Contrastive-LM (per the announcement: Stanford together with NVIDIA) · USA

An open fast-choice model for agents: it takes a description of a situation and a set of options and returns the probability of each without generating text. It can answer yes or no, pick from a list and give a score on a scale. Actions and state are encoded separately, so action vectors can be computed in advance: by the authors' measurements, with a choice among about a thousand options it is up to 13 times faster than the Jev analogue. After fine-tuning it works as a verifier that picks the best of several solutions of a coding agent (DeepSWE 81.6%, Terminal-Bench 2.1 87.6% by their measurements). It needs Linux, an NVIDIA card and Qwen3-8B in vLLM, and the card language is English.

  • fast choice of an agent's next action from given options
  • checking and ranking several solutions of a coding agent
  • routing requests: department, urgency, a score on a scale
Sizes
8B (a frozen Qwen3-8B plus two small projection heads)
Hardware
from: 1 GPU
Commercial use allowedDetails
Text analysis2026

Hy-Memory

Tencent Hunyuan · China

A model that runs the long-term memory of an AI agent: it decides what from a chat, work tasks or code is worth remembering, in what form, and how new information changes what is already stored. It replaces the general-purpose model inside the memory systems Mem0, MemOS and Tencent's own Hy-Memory. Languages on the card: English and Chinese. The benchmark figures are Tencent's own.

  • an assistant's memory about a customer between conversations
  • remembering the course of work on a project and code
  • updating stored facts when new data arrives instead of duplicates
Sizes
30B-A3B
Hardware
from: 1 GPU
Commercial use allowedDetails
Text analysis2026

JEV и GEV

AutoTrust AI · Singapore

Open decision models: for yes/no, pick-one and rate-0-to-5 questions they answer with probabilities in one pass, and can also reason when needed. Not affiliated with the closed Jev from TypeSafe.

  • routing requests to the right team
  • checking an agent action before it runs
  • rating urgency and tone of messages
Sizes
9B – 35B
Hardware
from: 1 GPU
Commercial use allowedDetails
Text analysis2026

pplx-decider

Perplexity · USA

A decision model built on Qwen3.8-27B: it picks an option, answers yes or no or gives a rating, and returns calibrated probabilities. It also understands images.

  • routing requests to teams
  • checking AI answers against sources
  • rating tone and urgency of emails
Sizes
27B
Hardware
from: 1 GPU
Commercial use allowedDetails

Tabular data 3

Tabular data2026

TabFM

Google Research · USA

Google's large tabular model: classification and regression from examples without training, with numeric and categorical columns. Weights are for non-commercial use only.

  • Pilot comparison with current scoring models
  • Exploring customer data
  • Testing churn hypotheses
Sizes
about 1.6B
Hardware
from: 1 GPU
Non-commercial onlyDetails
Tabular dataGGUF2024–2025

TableGPT2 / TableGPT-R1

Zhejiang University · China

A family for working with tables and databases: it understands data structure, writes parsing code and answers questions about exports.

  • Answering questions about tables and data exports
  • Automated data analysis with generated code
  • A helper for BI and internal reporting
Sizes
7B – 72B
Hardware
from: 1 GPU
Commercial use allowedDetails
Tabular dataNot maintained2024

TabuLa-8B

ML Foundations · USA

A foundation model for predictions on tables: it classifies and forecasts from a handful of examples, with no separate task-specific training.

  • Classifying table rows from a few examples
  • Predicting a value from a data row
  • Quickly testing hypotheses on new datasets
Sizes
8B
Hardware
from: 1 GPU
Commercial use with conditionsDetails

Faces 3

Faces2025

InfiniteYou

ByteDance · China

FLUX-based image generation that preserves a face: follows the prompt better and less often pastes the face like a sticker. Research-only license.

  • Portraits from one photo with a precise scene description
  • Testing characters for advertising
  • Comparing face-preservation methods
Sizes
adapter for FLUX.1-dev
Hardware
from: 1 GPU
Non-commercial onlyDetails
Faces2024

PuLID

ByteDance · China

Preserves a person's face when generating images from one photo, with less damage to style and background. Versions exist for SDXL and FLUX; the latter runs on a 16 GB card.

  • Portraits and avatars from one photo
  • Ad characters with a recognizable face
  • Photoshoot prototypes
Sizes
adapters for SDXL and FLUX.1-dev
Hardware
from: 1 GPU
Commercial use with conditionsDetails
FacesNot maintained2024

InstantID

InstantX (Xiaohongshu) · China

Generates images with a specific person's face from a single photo, without fine-tuning. Popular in ComfyUI, but the weights are for research only.

  • Portraits in different styles from one photo
  • Avatar and character sketches
  • Photoshoot prototypes
Sizes
adapter for SDXL
Hardware
from: 1 GPU
Non-commercial onlyDetails

Deepfake detection 3

Deepfake detection2024–2025

AIDE

Xiaohongshu, USTC and Shanghai Jiao Tong University · China

An AI-image detector made of several experts: some look at visual artifacts, others at noise. The hard Chameleon benchmark was released with it. It errs in both directions - a human reviews the result.

  • Checking realistic AI images without obvious artifacts
  • Comparing detectors on hard examples
  • Fine-tuning for your own type of content
Sizes
several experts based on ConvNeXt and CLIP
Hardware
from: 1 GPU
Commercial use with conditionsDetails
Deepfake detection2022–2025

TruFor

GRIP, University Federico II of Naples · Italy

Finds traces of editing and shows on a map which regions of an image look altered: suitable for scans of contracts, certificates and photos of documents. It errs in both directions - a person decides.

  • Checking scans of certificates and contracts for edits
  • Highlighting altered photo regions for an expert
  • Filtering out obviously redrawn documents before manual review
Sizes
a transformer model producing a map of suspicious regions
Hardware
from: 1 GPU
Non-commercial onlyDetails
Deepfake detectionNot maintained2023–2024

IML-ViT

Sichuan University and co-authors · China

An open model for finding forgeries in images: it outputs a pixel-level mask of altered regions. It errs in both directions; its map is a hint for an expert, not proof of a forgery.

  • Finding pasted and erased fragments in photos
  • Checking document scans for edits
  • A baseline when comparing manipulation-localization models
Sizes
a Vision Transformer based model
Hardware
from: 1 GPU
Commercial use allowedDetails

Fact-checking and judges 3

Fact-checking and judges2025

Atla Selene

Atla · UK

An 8B judge model: it scores another model answer against your criteria and writes a rationale. The judge itself makes mistakes and does not replace manual review on important tasks.

  • Scoring chatbot answers against your own criteria
  • Comparing two versions of a prompt or model
  • Filtering out weak answers before they reach a person
Sizes
8B
Hardware
from: 1 GPU
Commercial use with conditionsDetails
Fact-checking and judgesNot maintained2024

Patronus Lynx

Patronus AI · USA

Checks whether a chatbot invented a fact that is not in the source documents. The license is non-commercial. The checking model itself makes mistakes and does not replace manual review on important tasks.

  • Finding invented facts in AI assistant answers
  • Checking that answers rest on the attached documents
  • Filtering out answers before they go to a customer
Sizes
8B and 70B
Hardware
from: 1 GPU
Non-commercial onlyDetails
Fact-checking and judgesNot maintained2024

ArmoRM (RLHFlow)

RLHFlow · USA

An answer scorer that returns a breakdown across several attributes rather than a single overall score. The scorer itself makes mistakes and does not replace manual review on important tasks.

  • Choosing the best of several candidate answers
  • Preparing data for model fine-tuning
  • Scoring assistant answers across several attributes
Sizes
8B
Hardware
from: 1 GPU
Commercial use with conditionsDetails

Recommenders 3

Recommenders2024–2025

HLLM

ByteDance · China

A two-level design: one language model reads the item description, the second builds the shopper interest from history. Ready checkpoints at 1B and 7B are published, but they need fine-tuning for your own catalogue.

  • Recommendations from browsing and purchase history
  • Understanding an item from its title and description
  • Suggesting cards for new items with no statistics
Sizes
1B, 7B
Hardware
from: 1 GPU
Commercial use with conditionsDetails
RecommendersNot maintained2024

RecGPT

VinAI Research · Vietnam

A 7B language model further trained on review and product-card text for recommendation and rating-prediction tasks. The licence forbids commercial use; a store still needs its own order history.

  • Predicting the rating a user would give an item
  • Suggesting the next purchase from history
  • Researching text-based recommendations
Sizes
7B
Hardware
from: 1 GPU
Non-commercial onlyDetails
RecommendersNot maintained2023

LC-Rec

Renmin University of China · China

An approach where every item gets a short identifier code that a language model understands, and the model predicts that code. Delta weights for three categories are published; no licence is stated on the card.

  • Suggesting the next item from history
  • Turning a catalogue into short codes for the model
  • Recommendations from a shopper text request
Sizes
about 7B (a delta to LLaMA weights)
Hardware
from: 1 GPU
Commercial use with conditionsDetails

Math and reasoning 2

Math and reasoning2026

Limite (Violetto и Value Model)

Paradigma · not disclosed

A 1B model that does exactly one thing: it solves maths problems. It answers in a single turn, with no dialogue, and handles non-maths questions poorly. A companion Value Model scores someone else's solution step by step.

  • Solving maths problems in a single answer
  • Picking the best solution out of several candidates
  • Scoring reasoning step by step when training another model
Sizes
1B
Hardware
from: 1 GPU
Commercial use allowedDetails
Math and reasoningOllama2024–2025

QwQ

Qwen (Alibaba) · China

Qwen's first open reasoning model: it thinks step by step before answering and comes close to DeepSeek-R1 on maths tasks with only 32B parameters.

  • Calculations and formula checks
  • Complex analytics with step-by-step breakdowns
  • Checking the logic of contracts and internal policies
Sizes
32B
Hardware
from: 1 GPU
Commercial use allowedDetails

Moderation and safety 2

Moderation and safetyOllama2025

gpt-oss-safeguard

OpenAI · USA

Moderation by your own rules: you write the policy in plain text, and the model reasons and gives a decision with an explanation. Built on gpt-oss.

  • Moderation by internal company rules
  • Labeling disputed messages with an explanation
  • Checking reviews and listings before publishing
Sizes
20B – 120B
Hardware
from: 1 GPU
Commercial use allowedDetails
Moderation and safetyGGUFNot maintained2024

WildGuard

Ai2 (Allen Institute for AI) · USA

An open Ai2 filter: in a single pass it determines whether a request is harmful, whether a reply is harmful, and whether the bot refused needlessly. Works in English.

  • Checking requests to the bot
  • Checking bot replies
  • Finding unnecessary bot refusals on harmless questions
Sizes
7B
Hardware
from: 1 GPU
Commercial use allowedDetails

Voice assistants 2

Voice assistantsGGUF2026

PersonaPlex

NVIDIA · USA

A voice conversation partner based on Moshi that listens and speaks at the same time and can be interrupted. The role is set by text, the voice by a sample recording. English only.

  • A voice assistant with a set role
  • A conversation simulator for staff training
  • Voice interfaces without delay
Sizes
7B
Hardware
from: 1 GPU
Commercial use with conditionsDetails
Voice assistants2025

Kimi-Audio

Moonshot AI · China

A general-purpose audio model: speech recognition, answering questions about sounds, detecting emotions and voice dialogue. Trained on 13 million hours of audio; languages are English and Chinese.

  • Speech recognition
  • Detecting emotions and sound events
  • Speech-to-speech voice dialogue
Sizes
7B
Hardware
from: 1 GPU
Commercial use allowedDetails

Autonomous driving 2

Autonomous driving2025–2026

NVIDIA Alpamayo

NVIDIA · USA

Vision-language-action models for self-driving vehicles: they plan a trajectory from camera video and explain the decision in text. Used to develop and test autopilot systems, not as a ready-made autopilot.

  • Auto-labeling camera recordings to train your own driver assistance systems
  • Analyzing complex road scenes with text explanations
  • Testing autopilot systems in simulation on rare scenarios
Sizes
10B – 34B
Hardware
from: 1 GPU
Commercial use with conditionsDetails
Autonomous driving2026

Qwen-Drive

Alibaba (Qwen team) · China

An autonomous driving model based on Qwen3.5-4B: 3D detection of objects around the vehicle, answers to questions about the road scene and trajectory planning in one model.

  • A perception and planning prototype for autonomous vehicles on closed sites
  • Answering questions about camera recordings when reviewing incidents
  • Labeling road scenes to train your own models
Sizes
4B
Hardware
from: 1 GPU
Commercial use allowedDetails

Cybersecurity 2

Cybersecurity2022–2025

NVIDIA Morpheus (набор моделей)

NVIDIA · USA

An open NVIDIA bundle: a card-transaction graph plus a classifier, a profile of normal user behaviour, log parsing and root-cause search. A human makes the decision on a disputed operation; automatic blocking without review is not acceptable.

  • Highlighting suspicious card operations taking participant links into account
  • Tracking deviations of an account from its usual behaviour
  • Parsing logs and finding the root cause of an incident
Sizes
from a few megabytes (tree-based models) up to BERT-based classifiers
Hardware
from: 1 GPU
Commercial use allowedDetails
CybersecurityGGUF2025

Trendyol Cybersecurity LLM

Trendyol · Turkey

Security models from a large Turkish marketplace, published in GGUF format: reviewing alerts and incidents, English and Turkish.

  • Reviewing alerts and first-pass incident assessment
  • Explaining suspicious activity in reports
  • Helping the on-duty shift of a monitoring centre
Sizes
32B и 70B
Hardware
from: 1 GPU
Commercial use allowedDetails

Text to speech 1

Text to speechRUGGUF2025–2026

Higgs Audio

Boson AI · USA

Expressive speech and dialogue synthesis with voice cloning, plus recognition models. Version 3 of the synthesis supports about 100 languages, including Russian, but is non-commercial.

  • Expressive video voiceover
  • Voicing dialogues
  • Voice cloning
Sizes
about 3B – 8B
Hardware
from: 1 GPU
Commercial use with conditionsDetails

Translation 1

TranslationRUNot maintained2023–2024

ALMA и X-ALMA

Johns Hopkins University and Microsoft · USA

Research translators based on Llama 2. The first ALMA covered 5 pairs with English, including Russian; X-ALMA expanded coverage to 50 languages.

  • Translation between English and Russian
  • Experiments with LLM-based translation
  • Base for fine-tuning a translator
Sizes
7B – 13B
Hardware
from: 1 GPU
Commercial use with conditionsDetails

Voice: speakers and sound 1

Voice: speakers and sound2025

SAM Audio

Meta · USA

A model that cuts the sound you need out of a recording based on a text description, a mark on the video or a time range: a voice, an instrument, noise. Weights are available on request.

  • Isolating one person's voice from a noisy recording
  • Removing unwanted sound from a video
  • Splitting a recording into separate sound sources
Sizes
small, base, large (5 to 15 GB of weights)
Hardware
from: 1 GPU
Commercial use with conditionsDetails

Rerankers 1

RerankersGGUFNot maintained2023–2024

RankVicuna / RankLLaMA / RankZephyr (Castorini)

University of Waterloo, Castorini group · Canada

Rerankers that are language models: they receive the whole list of retrieved passages and reorder it as a list, instead of scoring passages one by one. Heavier than ordinary rerankers.

  • Reordering a long list of search results
  • Selecting sources for an AI assistant answer
  • Research comparisons of retrieval approaches
Sizes
7B – 13B
Hardware
from: 1 GPU
Commercial use with conditionsDetails

Plates and vehicles 1

Plates and vehicles2024–2025

VehiclePaliGemma

NYU Abu Dhabi (ComNets) · UAE

A model that looks at a photo and answers in words what the plate says: fine-tuned on Malaysian plates, so the Russian format needs fine-tuning. Filming vehicles and people is personal data processing: agree the procedure with a lawyer.

  • Reading plates in difficult frames
  • Answering questions about a vehicle photo
  • Going through an archive of camera photos
Sizes
3B
Hardware
from: 1 GPU
Commercial use with conditionsDetails

Speech to text 1

Speech to text2026

Audio8 ASR Infinite

Edge0 · not stated

Streaming recognition of Chinese and English speech with a 240-560 ms delay. It can listen indefinitely without memory growth and tells a thinking pause from the end of a turn.

  • subtitles for live streams and calls
  • voice assistants with fast reaction
  • round-the-clock transcription of a speech stream
Sizes
4B
Hardware
from: 1 GPU
Commercial use allowedDetails

Where people usually hit the wall

Plan by video memory, not by parameter count: the weights, the context and the request queue all consume it. A model that fits the weights exactly will fail on a long document. Leave headroom and test on your own scenario rather than a demo prompt.

By hardware

Need a model for your task?

An open model can run on your own server: data stays in-house, there is no per-request fee, and the model can be fine-tuned on your documents.

  1. SelectThe model and size for your task and hardware budget
  2. DeployOn your server or in a closed network, with an API
  3. Fine-tuneOn your data, or connect a knowledge base
  4. IntegrateInto your CRM, ERP, bot, website or team chat
Discuss deployment