Without a GPU an open model runs on the CPU and ordinary RAM. It is slower, but you buy nothing extra: a good fit where the answer is not needed instantly and volumes are modest — overnight document processing, transcription, knowledge-base search.
DeepSeek's flagship line: from the first 7B/67B to V4-Pro with 1.6 trillion parameters. Closed-model quality under an open MIT license; V4-Flash-Vision-Exp and V4.1-Flash understand images, context up to 1M tokens.
Models from Shanghai AI Laboratory. The early InternLM line is general-purpose; the new Intern-S1/S2 is scientific: it understands formulas, molecules, charts and images.
Xiaomi models for reasoning and agents: from the compact MiMo-7B to MiMo-V2.6-Pro with 1.02 trillion parameters. The larger versions understand text, images, video and audio, with a 1M token context. Languages: English and Chinese.
Sber open models with strong Russian language support and local context, from 10B-A1.8B to 702B, all MIT. GigaChat3.1-Audio handles recordings up to two hours; GFusion is a fast diffusion text version.
Yandex models trained from scratch with a focus on the Russian language and Russian context. The new AliceAI-Foundation 80B-A3B (Apache 2.0) is a base model only, with no instruct version: you fine-tune it for your own tasks. The efficient AliceAI-T5 35B-A0.6B is also available.
Multilingual models from Cohere's research arm, covering 23 to 100+ languages. Tiny Aya (2026, 3.3B) runs on a regular PC, but for non-commercial use only.
Compact text models that run directly on a device: laptop, phone or mini PC. The 1B and 2B MiniCPM5 models focus on tool calling and long context.
Fully open models from the UAE: data, training code and intermediate checkpoints are published along with the weights. K2-Horizon (2026) spans 0.9B to 375B with context up to 512K tokens.
A Russian and English research prototype: the model writes text in blocks at once (diffusion) rather than word by word, which speeds up responses. The authors do not recommend it for production systems.
A family of language models with strong Russian language support, from small versions for a laptop to a flagship on par with commercial APIs.
One of the oldest Chinese open lines: from ChatGLM-6B to GLM-5.3. Strong at agentic tasks and programming; GLM-5.3-Flash understands images and is released under MIT.
Tencent language models: from small 0.5B–7B to Hy4-preview with 770 billion parameters. Since 2026 the line has been renamed Hy, and new versions are released under Apache 2.0.
Speech synthesis with voice cloning and precise duration control, handy for video dubbing. Controls emotion separately from timbre.
Speech synthesis with voice cloning and fine control over emotion, speed and pitch.
Speech synthesis with voice cloning and adjustable expressiveness. The multilingual version supports 23 languages, including Russian; Turbo and Flash are sped up for live dialogue.
A speech synthesis family: multi-voice dialogue voicing (TTSD), fast synthesis for live conversation and the tiny Nano. Version 1.5 supports 30+ languages, including Russian.
Very fast, lightweight speech synthesis that runs directly on the device, without a GPU or the cloud. Supertonic 3 speaks 31 languages, including Russian.
Speech synthesis with voice cloning and natural intonation. VoxCPM2 supports 30 languages, including Russian.
Streaming speech recognition and synthesis models from the makers of Moshi: they start speaking and transcribing without waiting for the end of a phrase. Pocket TTS (100M) runs on a CPU. English, French and a few other European languages, no Russian.
Speech synthesis with voice cloning and emotion control in 80+ languages, including Russian. Quality is close to paid services, but the weights are for research only.
Speech synthesis with voice cloning from a short sample in 646 languages, including Russian and languages of Russia's peoples. A voice can be described in words. Weights are for non-commercial use only.
Speech synthesis in 10 languages, including Russian: voice cloning from 3 seconds, ready-made voices and creating a voice from a text description.
Speech synthesis with voice cloning from a short sample and streaming output for live dialogue. Version 3 supports 9 languages, including Russian.
Lightweight Russian speech synthesis that runs on a regular CPU. Version v5 added CIS languages and languages of Russia's peoples: Tatar, Bashkir, Yakut, Kazakh and others.
A new lightweight document parsing model that led the OmniDocBench v1.6 benchmark at release. Handles pages photographed on a phone and crumpled pages well. Languages on the card: Chinese, English, Japanese.
A lightweight OCR model from Tencent: document parsing, finding text in photos, field extraction and translating text from images. Version 1.5 is faster and runs on an ordinary PC.
A compact OCR toolkit from the makers of Marker and Chandra: text recognition, page layout, reading order and tables. Surya OCR 2 (650M) also runs on a CPU; Russian scored 88.8% in benchmarks.
Baidu's OCR model building on DeepSeek-OCR ideas: processes multi-page documents and PDFs in a single pass and outputs structured text. Claimed to be multilingual, but the language list is not published.
Classic lightweight PaddleOCR models: detecting and recognizing lines of text plus page layout. They run on CPUs and phones; there is a separate model for East Slavic languages, including Russian.
A compact document parsing model from the popular PaddleOCR toolkit. Per the model card it supports 109 languages, including Russian; version 1.6 leads the OmniDocBench benchmark.
A popular open tool for converting PDFs to Markdown with its own small model. MinerU2.5-Pro was improved through data alone, without growing in size. Languages on the card: Chinese and English.
A multilingual document parsing model: text, tables, formulas and reading order in one pass. dots.mocr also turns charts and diagrams into vector SVG.
A strong OCR model from the authors of Marker and Surya: handwriting, forms, tables. Per the model card it supports 90+ languages, with Russian among the examples.
A Baidu model that not only recognises a document but also answers questions about it. Per the model card it supports 192 languages, including Cyrillic.
An OCR model that compresses a page into a small number of visual tokens, so it processes large volumes quickly. Version 2 better understands reading order.
A lightweight OCR model from Zhipu for document parsing. The model card lists Russian among supported languages; built for high load and low-end hardware.
Sber embeddings built for Russian: according to the developers, among the best on Russian-language search benchmarks. FRIDA is compact, Giga-Embeddings is more powerful.
NVIDIA embeddings for search and RAG. Nemotron-3-Embed, released in 2026, is under the permissive OpenMDW license and works in many languages.
Strong multilingual embeddings with long context; v5-omni understands text, images and audio. Recent versions are open for non-commercial use only.
Lightweight IBM embeddings for enterprise search, trained on data with clear rights. R2, released in 2026, became multilingual.
Qwen3-Embedding models fine-tuned by the startup Octen for search in legal, financial and medical texts. As of January 2026 the 8B version topped the RTEB leaderboard.
Embeddings from the Perplexity search service. Some versions take into account the context of the whole document, not just a single fragment.
Microsoft's 2026 multilingual embeddings with context up to 32K tokens; Russian is on the language list. The 270M and 0.6B versions run on a regular server, 27B is the most accurate.
Embeddings and rerankers based on Qwen3, among the best open ones for multilingual search, including Russian. VL versions search images, screenshots and video.
The only open model in the Voyage 4 line: its vectors are compatible with the paid larger versions, so you can start locally and move to the API later.
Embeddings and rerankers from Germany's Mixedbread. mxbai-embed-large is one of the most downloaded English search models; the v2 rerankers cover 100+ languages, including Russian.
A small multilingual embedding model based on Gemma 3 that runs even on a phone or laptop without internet.
Some of the most popular embeddings for search and RAG. The main v1.5 versions target English and Chinese; for Russian, BAAI has a separate model, bge-m3.
Vision models from Alibaba's international division with strong text and table reading. The line includes the Ovis2.6 MoE and separate compact OvisOCR models for documents.
A small, fast vision model for product use cases: answering questions, finding and pointing to objects, captions. Moondream 3.1 is a 9B MoE with 2B active.
A family of video models: encoders for search and classification of clips, and chat models that analyze long videos. InternVideo 3 is designed for multi-hour recordings.
The open project that started the trend for image-plus-text models. The OneVision line understands photos, documents and video; training data and recipes are open.
Compact vision models that run even on a phone or laptop. Good at reading text in photos and understanding video; version 4.6 is only 1.3B.
Vision models from Kuaishou focused on short videos. Keye-VL-2.0 (30B, 3B active) understands well what happens in a clip and when.
Compact IBM models for business documents: tables, charts, forms, field-value pairs. The model card openly warns that it works best with English.
A large family of Chinese vision models sized from 1B to 241B. InternVL-U (4B) combines image understanding, generation and editing.
Fully open vision models from Ai2 (weights and data). They can point to a spot in an image and count objects; Molmo2 understands video, MolmoWeb controls a browser.
Vision models from Zhipu: first CogVLM, then the GLM-V line. GLM-4.6V can call tools based on images and act as an agent operating an interface.
One of the strongest open vision models: reads documents, tables, charts and video, and works with user interfaces. Since Qwen3.5, vision is built directly into the main Qwen model.
The smallest vision models from Hugging Face, starting at 256M; they run in a browser and on a phone. SmolVLM2 also understands video.
A large open collection of models for separating vocals from music and noise: MDX-Net, BS-RoFormer, Mel-RoFormer, SCNet. The quality leaders for vocals among open solutions.
A set of ready-made voiceprint models: checks whether the same person speaks in two recordings and helps split a recording by speaker. One of the models is built into pyannote 3.x.
A classic model that splits a track into vocals, drums, bass and the rest. The v4 hybrid transformer version remains the benchmark; the project is now maintained by its author in his own repository.
The most widely used open voice conversion tool: a model for a specific voice trains on 10–30 minutes of recording and works in real time. Use only with the voice owner's consent.
A speech and sound event detector: tells apart speech, singing and music. In a 102-language test (the FLEURS set, which includes Russian) it beat Silero VAD and TEN VAD. Has a streaming mode.
Uses intonation to tell whether a person has finished a thought or just paused, so a voice bot does not interrupt. Version 3 is 8 MB, runs on a CPU and understands 23 languages, including Russian.
The most popular open speech detector: tells voice apart from silence and noise. Processes an audio chunk in under a millisecond on a single CPU core; trained on recordings in more than 6,000 languages.
NVIDIA models for "who is speaking": TitaNet recognizes a specific person's voice, Sortformer splits a recording into up to 4 speakers, including live during a call.
The most widely used open tool for splitting a recording by speaker: who spoke and when. Usually paired with speech recognition. Weights are issued after a short form on HF.
Alibaba's set of speech cleanup models: noise suppression, separating overlapping voices, upscaling audio to 48 kHz, and isolating a voice using video of the speaker's face.
A lightweight speech detector for real-time voice assistants: it notices the start and end of a phrase faster than Silero VAD. Runs on servers, phones and in the browser.
Models that detect emotion from voice in Russian speech: neutral, anger, positive, sadness. Trained on the open Dusha dataset from SberDevices.
Reconstructs the 3D geometry of a scene from one photo: depth in meters, a point cloud and surface normals.
Real-time object detector, an open alternative to YOLO without AGPL. Supports segmentation (object outlines) and, since 2026, keypoints.
Meta's models for analyzing people in photos: pose keypoints, body part segmentation, normals and depth. Sapiens2 was trained at high resolution and adds human matting.
Selects any object in photos and videos with a click or a box. The basis for background removal and object counting.
The most widely used real-time object detector: finds and marks items in video even on modest hardware. YOLOv5 came out back in 2020; the catalog starts from YOLOv8.
Estimates depth, the distance to every point, from one ordinary photo or video. DA3 reconstructs scene geometry from several frames.
Meta's family of encoders for images and video, and with PE-AV also for audio. PE-Core searches by text more accurately than SigLIP 2 (per Meta); small versions are available.
Meta's open reproduction of CLIP with a transparent data collection recipe. MetaCLIP 2 is trained on multilingual data from around the world. Non-commercial license only.
Finds any objects in an image from a text description, without training on your data: "red box", "person without a hard hat". Rex-Omni is the new VLM-based generation.
Foundation models that turn an image into a numeric "fingerprint". They are used to build similar-image search, classification and segmentation without large labeled datasets.
Models that map images and text into a shared space: you can search photos by words and classify images without training. OpenAI's CLIP (2021) is the predecessor.
A simple, accurate model for human pose estimation via keypoints. ViTPose++ handles human, animal and whole-body poses; built into the Transformers library.
A large family of medical models: chat, an imaging version, the reasoning HuatuoGPT-o1 and the new HuatuoGPT-3 on Qwen3. Does not replace a doctor; decisions are made by a specialist.
Open medical models from Swiss EPFL, fine-tuned on clinical guidelines on top of various base models. Does not replace a doctor; decisions are made by a specialist.
A medical model for text, images, 3D scans and video: from a light 4B to a large MoE. Does not replace a doctor; decisions are made by a specialist.
Google's medical version of Gemma: reads medical texts and images (X-ray, dermatology, histology). A tool for doctors and developers; does not replace a doctor, decisions are made by a specialist.
Stanford models for chest X-rays: they describe the image and prepare a draft report. Does not replace a doctor; decisions are made by a specialist.
Alibaba's medical model based on Qwen2.5-VL: understands many types of medical images and medical text, and can reason step by step. Does not replace a doctor; decisions are made by a specialist.
Google's lightweight encoder for medical images and text, the same one inside MedGemma. Sorts images and finds similar ones. Does not replace a doctor; decisions are made by a specialist.
Reasoning medical models on Qwen3, designed to run on an ordinary computer. Does not replace a doctor; decisions are made by a specialist.
Pathology foundation models from France's Bioptimus with 1.1B parameters, plus the compact H0-mini. The first version is open under Apache 2.0. Does not replace a doctor; decisions are made by a specialist.
A foundation model for histology slides: turns patches of digital slides into features for tissue classification. Does not replace a doctor; decisions are made by a specialist.
Image-plus-text models for pathology: search slides by an English description, classify without fine-tuning; TITAN describes a whole slide. Does not replace a doctor; decisions are made by a specialist.
Paige's pathology foundation model, trained on millions of digital slides. The first version is Apache 2.0, the second is for research only. Does not replace a doctor; decisions are made by a specialist.
An image watermark for arbitrary resolutions built for the Content Authenticity Initiative: it can both apply a mark and remove one. The detector errs in both directions - a human reviews the output.
A lightweight detector of generated images, trained on 2.7M samples from nearly 5000 different generators. It errs in both directions: the result is a reason for a human to check, not proof.
An early and still used approach: a simple classifier trained on top of a frozen CLIP that transfers to unseen generators. It errs in both directions - the output needs a human check.
A watermark for video and images that survives re-encoding and cropping. The detector errs in both directions: a missing mark does not prove a forgery, and finding one is a reason for a human to check.
Seven speech encoders (wav2vec 2.0, XLS-R, MMS, HuBERT) post-trained to tell live speech from synthetic. The authors note themselves that quality depends heavily on the dataset; a human reviews the output.
An AI-text detector trained together with a paraphraser: it was deliberately taught not to give up when the text has been rewritten. It errs in both directions; a human reviews the output.
Dozens of open face-swap detectors for video and photo under one codebase with ready weights. A detector errs in both directions: its output is a reason for a human to check, not proof of a forgery.
An image watermark that can be applied to individual regions: the model shows which part of the image is marked. It errs in both directions - a human reviews the result.
A recent open AI-text detector on DeBERTa-v3-large, trained on the RAID dataset, with a separate version for academic work. It errs in both directions - a human always reviews the result.
An imperceptible mark in synthetic speech plus a fast detector that finds it even inside a fragment of a long recording. The detector errs in both directions: a hit is a reason for a human to check, not proof.
A Longformer-based AI-text detector: it holds a long document whole and was trained on texts from many different language models. It errs in both directions; its output is a reason for a human to check.
One of the first open AI-text classifiers, trained on the HC3 corpus of paired human and ChatGPT answers. It errs in both directions: its output is a reason to talk to the author, not proof.
JetBrains models for fast code autocompletion. Mellum2 (12B, 2.5B active) is already a full assistant: it writes and edits code, calls tools and reasons.
An 8B model trained from scratch by the company of one of the authors of the transformer architecture. Strong at code and technical tasks; version 1.5 handles context up to 160K tokens.
The broadest open coding family: from 0.5B for autocompletion to 480B for agents. Qwen3-Coder-Next (80B, 3B active) works as a developer agent on a single GPU.
A family of coding models with standard and reasoning versions, including a Loop variant that runs through its layers a second time. Sizes from 7B to 40B.
Fully open developer agents from Ai2: weights, data and training recipe are all public. Designed so a company can cheaply fine-tune the agent on its own repository.
A compact 8B coding model from ByteDance in base, instruct and reasoning versions. Its training data was selected by the model itself, with almost no hand-written rules.
A small coding assistant from MTS AI that understands requests in Russian. Runs locally, with plugins for VS Code and JetBrains.
Fully reproducible coding models: along with the weights, the data, its cleaning pipeline and the training recipe are open. Understand English and Chinese.
DeepSeek's coding model family: from small autocompletion models to the large MoE V2, which matched closed models in 2024. Later, coding moved into DeepSeek's general models.
Coding models from 01.AI at 1.5B and 9B with a 128K-token context and support for 52 programming languages. A separate line next to the text Yi models.
Mistral's coding model covering 80+ programming languages. The open weights of the main version cannot be used in production without a paid license; newer Codestral versions are API-only.
Coding models from the creators of GLM. CodeGeeX4-ALL-9B, based on GLM-4-9B, combines autocompletion, code chat, function calling and repository search in one model.
BRIA's background removal, trained on licensed photos. Soft edges, hair, transparency. Video versions available. Business use requires a paid agreement.
Cuts a person out of video with a precise alpha mask, including hair and edges, without a green screen. Needs a first-frame mask, for example from SAM.
An open MIT-licensed model for precise object segmentation and background removal. RMBG-2.0 is built on it. Versions for 2K and for hair and semi-transparent edges.
A background removal model focused on difficult edges: hair, fur, fine details. The open version is MIT-licensed and can process video.
All-round photo inpainting: remove an object, insert a new one from a description, change a shape or extend the frame beyond its edges.
Changes lighting in a photo: relights an object or person from a description or to match a given background, so a cut-out looks natural.
Colorizes black-and-white photos in natural colors. A lightweight model with a commercial-friendly license; a compact tiny version is available.
A salient object detection model and the ready-made transparent-background tool built on it: removes backgrounds from photos, video and webcam with one command.
Transformer-based photo upscaling, more accurate than SwinIR on fine details. Versions for real noisy photos and a lightweight HAT-S.
Popular face restoration for old and blurry photos; also works on video. Has face inpainting and colorization modes. Non-commercial license.
The classic for upscaling photos 2–4x while cleaning noise and compression artifacts. Lightweight, runs even on a CPU. Versions for drawings and anime.
Proven face restoration for old and compressed photos, with a commercial-friendly license. Often paired with Real-ESRGAN; the most used versions are 1.3 and 1.4.
Microsoft speech models: long multi-voice dialogue synthesis, fast synthesis for live conversation, and recognition of long recordings split by speaker, including in Russian.
Very small and fast speech recognition models for phones, tablets and embedded devices. Version 2 streams, producing text while the person is still speaking.
IBM speech models for recognizing and translating speech in English, several European languages and Japanese. Designed for enterprise use.
Sber's models for Russian speech recognition, among the most accurate for Russian. Includes emotion recognition, v3 with punctuation, and a multilingual version (Russian, Kazakh, Kyrgyz, Uzbek).
Speech recognition models from the Qwen team for 50+ languages, including Russian. They handle noise, singing and accents well.
Cohere's speech recognition model for 14 languages (Russian is not on the list), with a separate version for Arabic. Built for accurate transcription of business recordings.
Fast NVIDIA speech recognition models, including streaming ones for real-time use. Parakeet TDT v3 and Nemotron 3.5 ASR understand Russian.
Mistral's speech models: they understand audio, transcribe and answer questions about a recording. The Realtime version recognizes speech live and supports Russian; speech synthesis is also available.
Speech recognition for 1,600+ languages, including Russian and rare languages no system supported before. A new language can be added from a few examples.
Alibaba's set of fast speech recognition models, primarily for Chinese and Asian languages. SenseVoice also detects emotions and sound events.
Offline Russian speech recognition that runs even on a Raspberry Pi or a phone, without internet. Streaming models for live audio and simple Russian speech synthesis, Vosk TTS, are available.
A compact T-Bank streaming model for recognizing Russian speech in phone calls. Works in real time even without a GPU.
Finds the entities you need in text without training: just list what to look for (name, amount, date). GLiNER2 also classifies text. Multilingual versions understand Russian.
Models for template-based data extraction: give it a document or scan and a JSON field template, get a filled-in JSON back. NuExtract3 (4B) also converts scans to Markdown.
Russian encoders from the VK team: RuModernBERT reads long texts, USER produces vectors for search, GeRaCl classifies texts by topic without training.
A modern replacement for classic BERT: faster, reads up to 8 thousand tokens at once. A base for your own classifiers. Trained on English and code; for Russian there is RuModernBERT.
Small models that turn raw web page HTML into clean Markdown or JSON. Handy for preparing websites for a knowledge base. Non-commercial license only.
Sber's Russian-language encoders trained on large Russian corpora. A base for classifiers, NER and semantic search in Russian.
Turns a free-form occupation description into a standard HISCO code in 13 languages. Built for historical archives, but also useful for cleaning up job title reference lists. A human makes the decision about a candidate; automatic screening without review must not be used.
A research line of models for labour market texts: trained on job postings and the European ESCO occupation taxonomy, they pull skills and requirements out of vacancies. A human makes the decision about a candidate; automatic screening without review must not be used.
Ready-made tiny rubert-tiny models for Russian text: detect rudeness and insults, sentiment and emotions. They run on a CPU in milliseconds.
Classic document understanding models: they take into account the text, its position on the page and the image. They are fine-tuned to extract fields from forms and receipts. Only the first version is free for commercial use.
A classic multilingual encoder for 100 languages, including Russian. The base of many sentiment, NER and embedding models, including BGE-M3.
A time-tested encoder behind many classifiers and NER models (including GLiNER). The multilingual mDeBERTa-v3 understands Russian.
Fully open reasoning models: both weights and training data are published. Newer OpenThinkerAgent versions can carry out multi-step tasks.
Open models for formal proofs in Lean 4 from Princeton. The new Goedel-Code-Prover proves program correctness.
A small 4B model on Qwen3 that writes mathematical proofs in plain language almost at the level of large models. Runs on a laptop.
DeepSeek's maths models. The first 7B version introduced the GRPO training method; the 685B V2 writes and checks its own olympiad-level proofs.
Models for formal proofs in Lean 4 from Moonshot AI (Kimi) and Numina. Small versions from 0.6B run on a laptop.
Small models fine-tuned with reinforcement learning: DeepScaleR (1.5B) solves olympiad maths, DeepCoder writes code, DeepSWE works as a developer agent. Recipes and data are open.
NVIDIA models for maths and reasoning based on Qwen. AceReason was fine-tuned with reinforcement learning first on maths, then on code.
DeepSeek models for formal proofs in Lean 4: the proof is checked by a program, not a person. A narrow tool for mathematicians and engineers.
A reasoning model trained on just a thousand problems. It can be told to think longer to answer a hard question more accurately.
Reasoning models from Qihoo 360: a standard Qwen2.5 was fine-tuned for long reasoning using an open recipe; data and code are published.
A Berkeley reasoning model trained for under 450 dollars. It showed that o1-preview-level reasoning can be reproduced with modest resources.
Maths versions of Qwen: they solve problems step by step and can calculate via code. Includes reward models that check each step of a solution.
A ready-made Google forecasting model: forecasts any time series without training on your data.
A compact forecasting model on the xLSTM architecture, a leader in open benchmarks despite its small size. Runs fast on a regular CPU.
A compact forecasting foundation model from the Tsinghua lab: trained on a large set of diverse series and fine-tunable on your own data.
A Datadog forecasting model trained on server and application metrics. Especially strong for IT monitoring: load, latency, errors.
Amazon forecasting models, among the most downloaded. Chronos-2 takes external factors into account: prices, promotions, weather.
Salesforce's universal forecasting model for series with different frequencies and many variables. Weights are open for research only.
A forecasting model that returns a set of possible scenarios rather than a single line — useful when you need a range for demand or load, not one number.
A foundation model for numeric series: one engine is used for forecasting, anomaly detection, filling gaps and classification.
Tiny forecasting models from IBM: they run on an ordinary CPU and sit next to the business system without a separate GPU server.
A forecasting model with a sparse architecture: only part of the network runs at each step, so it stays fast at a small size.
One of the first open out-of-the-box forecasting models. Tiny, gives a probabilistic forecast, now behind Chronos and TimesFM.
Small GLiNER-based models for finding personal data: passports, phone numbers, accounts, addresses. Data types are set in words. Russian is not officially supported.
NVIDIA content filters for bots, with separate models for keeping the conversation on topic and detecting jailbreaks. Safety Guard v3 was trained on 9 languages; Russian was tested only without fine-tuning.
IBM judge models: they catch harm, profanity and jailbreak attempts, and in RAG and agents check whether an answer is grounded in the documents. You can state your own rule in words.
Finds and hides personal data: names, addresses, phone numbers, emails, account numbers, passwords. Runs even in the browser. Trained mostly on English.
A guard model that catches both harmful content and attacks on AI (prompt injection, jailbreaks), including when agents use tools.
Safety filters for 119 languages, Russian among them. The Stream version checks a bot's reply while it is being generated and can cut it off on the fly.
Filter models that check chatbot requests and replies for dangerous topics against a list of categories. Version 4 also checks images. Russian is not officially supported.
Tiny classifiers that catch attempts to hack a bot: prompt injections and rule bypassing. The 86M version is multilingual, 22M is English only.
Small models that tell explicit images from regular ones. The Freepik model distinguishes four levels of explicitness. They run on a CPU.
Gemma-based filters: they check text for dangerous and offensive content, and ShieldGemma 2 checks images. Focused on English.
A popular detector of 17 types of personal data in six European languages. No Russian and a non-commercial license: suitable for trials and research.
A text-to-SQL model of just 1B parameters, designed to run locally so the database never leaves for external services.
A text-to-SQL model trained with reinforcement learning: it works through the schema and the conditions step by step before producing a query.
A Snowflake model for turning questions into SQL, trained with reinforcement learning by checking query results. The open 7B version is based on Qwen2.5-Coder.
Alibaba models for turning questions into SQL, based on Qwen2.5-Coder. They work with different SQL dialects; a small 3B version suits modest hardware.
Models for turning questions into SQL, trained on millions of synthetic query examples across different databases. Three sizes for different hardware.
One of the first open models that turn a plain-language question into an SQL query against a database. Available in Ollama, but newer competitors are already stronger.
A text-to-SQL model from the open Chat2DB database client: it supports different SQL dialects, with an English and Chinese model card.
A text-to-SQL model built on DeepSeek-Coder, aimed at complex questions spanning several tables and conditions.
A model for turning questions into SQL, built for the embedded analytics database DuckDB. Available in Ollama, convenient for working with CSV and Parquet locally.
An early line of open text-to-SQL models starting at 1B, including variants fine-tuned for specific database schemas.
One of the first open text-to-SQL lines, including very small versions from 350M that run on an ordinary PC.
A music encoder: turns a track into a numeric representation used to detect genre, mood, key and rhythm. MERT-v2 handles full songs up to 6 minutes.
Fast generation of songs with vocals in 19 languages, including Russian: a full song in seconds, editing of individual parts and style changes.
Generates short music clips and sound effects from a description. Version 3 is split into separate models for music and for sounds.
Fast generation of a full song with vocals from lyrics and a style sample, up to several minutes long.
Adds sound to silent video: generates noises and sound effects in sync with the on-screen action, from the video and a text prompt. One of the first strong open Foley models.
The MuQ music encoder and the MuQ-MuLan model, which matches music and text: you can search for tracks by a description in English or Chinese.
Generates instrumental music from a text description or a sample melody. One of the first open models of its kind.
CLIP for audio: maps audio and text into a shared space. Lets you search sounds and music by description and classify them without training. Text must be in English.
A classic 2021 sound recognition model: detects 527 AudioSet event classes (siren, barking, breaking glass, music). Lightweight, runs without a GPU, in Transformers since 2022.
Tencent translators for 33 languages; the first version won the WMT25 competition. Russian is supported. The small 1.8B version runs on a laptop; the new Hy-MT2 is under Apache 2.0.
More than a thousand small translators, each for its own language pair. Russian-English and back are available. Fast even on a regular CPU.
Translators based on Gemma 3 for 55 languages that can also translate text in images. Russian is supported. The 4B version fits on a laptop.
A compact ByteDance translator for 28 languages, close in quality to large closed systems. Russian is supported. Ready-made compressed versions are available.
Language models tailored for translation and multilingual text work: translating, editing, and assessing translation quality. Russian is supported. Non-commercial license.
Compact Kazakh-Russian translators from VK. At 197M they translate as well as the 600M NLLB and run on a regular CPU.
Google's translator for more than 400 languages under a permissive license. Russian is supported. A good substitute for NLLB when commercial use is needed.
A translator for 200 languages, including rare and minor ones. Russian is supported. Strong language coverage, but the license prohibits commercial use.
An early Meta translator that translates directly between 100 languages, without English in the middle. Russian is supported. Old, but light and permissively licensed.
An Ant Group family for finding elements on screen and completing tasks in phone and computer interfaces. UI-Venus-2 was specifically trained to refuse dangerous actions.
Small Microsoft models for working in the browser: they look at the page and click, type and scroll. Designed to run directly on a work computer without the cloud.
A French model family for controlling a browser and computer: precisely finds the right element on screen and handles multi-step tasks. The latest Holo3 and 3.1 are open under Apache 2.0.
A lightweight model for working with interfaces: finds buttons and fields by description and performs actions on the web and on a phone. ShowUI-π can drag with the mouse.
Models for controlling phones and computers from the Mobile-Agent project: they work with Android, Windows, macOS and the browser; version 1.5 has a reasoning mode.
Compact Alibaba models for working in smartphone and computer interfaces: they find elements and complete multi-step tasks. The small size allows running on an ordinary GPU.
A model that looks at a screenshot and controls the mouse and keyboard itself: clicks, fills in fields, navigates menus. The first generation and 1.5-7B are open; UI-TARS-2 weights were not released.
Salesforce models for function calling and agents: they pick the right tool and fill in its parameters. Strong on benchmarks, but the license is non-commercial.
Breaks a screenshot down into buttons, fields and icons with labels so a regular language model can understand and control the screen. It does not click itself; it serves as the agent's eyes.
A ready-made model for tables: it takes example rows and immediately predicts for new ones, without lengthy training or tuning. Only v2 is free for business; newer versions are non-commercial.
Amazon's tabular model built into AutoGluon: classification and regression from examples with brief fine-tuning. Mitra-v2 handles more rows and columns.
A tabular model from a Canadian bank's AI lab, trained on real tables rather than only synthetic ones. Version 1.2 Turbo made computation orders of magnitude faster.
A table model that alone can classify, predict numbers and fill in missing data. The lightweight LimiX-2M runs on an ordinary computer.
LG's small tabular model: with 21M parameters it nearly matches the leaders in classification and regression accuracy. Weights are for non-commercial use only.
Recent open models for tabular data: they predict from a few examples given in the prompt, with no task-specific training.
An open tabular model from the creators of scikit-learn: classifies and predicts from examples without training and handles tables of up to hundreds of thousands of rows. The license allows business use.
A model for office work with tables: for a given question it returns either a direct answer or code to process the data in a table or document.
A general-purpose model for tables: filling gaps, finding rows, matching columns and answering questions about the data.
Judge models that evaluate other AI models' answers in Russian: they score against a given criterion and explain the score in text.
A line of judges from the team behind open model benchmarks: they score answers and check them against a reference. The judge itself makes mistakes and does not replace manual review on important tasks.
Reward models: they score how good a language model's answer is for the user. Used for fine-tuning your own models and picking the best of several answers.
A small judge: it scores against your criteria and highlights which part of the answer led to that score. The license is non-commercial. The judge itself makes mistakes and does not replace manual review.
A small judge model: it checks an answer against your instruction and gives a score with an explanation. Fits on a modest server. The judge itself makes mistakes and does not replace manual review.
Checks whether each claim in an AI answer is supported by the source documents. The small versions are free; the larger 7B is in Ollama but non-commercial.
A small model that checks whether an AI answer is grounded in the source text or made up. Runs on a CPU and works well as a filter in RAG systems.
An open judge model: it scores other models' answers against your criteria and explains the score. A replacement for paid models in the reviewer role.
Strong multilingual rerankers; m0 also ranks pages as images (scans, slides). The latest versions are open for non-commercial use only.
A pair of small Tencent models based on Qwen3 that pick the right skill for an AI agent for a given request: the embedding model finds candidates, the reranker chooses the best one.
A small 1B reranker from NVIDIA. The vl version also takes document pages as images, not just text. The card states multilingual support without listing the languages.
Rerankers built on Qwen3. The model card lists the target domains — finance, law, code, medicine, science; the stated language is English.
A different search principle: every word of the question is compared with every word of the document, not the two texts as a whole. The index is heavier than with ordinary embeddings. The model cards list English.
Rerankers: they take passages found by search and reorder them by how well they actually match the question. v2-m3 is multilingual and lightweight, often paired with bge-m3.
An embedding-plus-reranker pair for knowledge bases. The card lists English, Chinese, Japanese and Korean — Russian is not among the stated languages.
The most downloaded open rerankers: a tiny model reads a question-passage pair and scores how well they match. The multilingual mMARCO version covers Russian.
Cisco models for information security based on Llama 3.1 8B: analysis of vulnerabilities, threats and incidents. Can be deployed inside your own perimeter.
A Chinese open family for cybersecurity: reviewing vulnerabilities, analysing logs and traffic, explaining commands and scripts.
One of the best-known open families for security and DevSecOps work: reviewing code for weaknesses, test scenarios, explaining attacks.
Trend Micro cybersecurity models based on Llama 3.1 8B, fine-tuned on a corpus of security texts. A reasoning version is available.
A very light classifier for emails and links showing signs of phishing. It errs in both directions, so borderline emails are still reviewed by a person.
A small open assistant for security professionals: questions about standards, reviewing threats and vulnerabilities, drafting internal documents.
A compact language encoder trained on cybersecurity texts: tagging threat reports, finding entities and classification.
Tiny audio-understanding models for smartphones: they listen to speech, music and ambient sounds and answer in text - describing a recording and answering questions about it. They run on the device itself; prompts and answers are in English - no other languages are present in the training data.
Models that listen to speech, sounds and music and answer questions about them. Audio Flamingo Next handles recordings up to 30 minutes. Research use only.
A voice assistant that listens and speaks at the same time, with no delay for recognition and synthesis. Hibiki does simultaneous speech-to-speech translation between several European languages.
A small model that sees, hears and replies by voice in real time, and can clone a voice. Voice dialogue in English and Chinese, text in 30+ languages.
Models that understand text, images, audio and video and reply by voice in real time. Qwen3-Omni speaks 10 languages, including Russian.
Speech and text translation across roughly a hundred languages, including Russian: speech to text, speech to speech, and streaming translation that keeps intonation.
Matches lip movements in an existing video to a new voice track. Version 1.6 works at 512 pixels and produces a sharper face.
Real-time lip sync: matches the mouth in a video to new audio. Suits video translation and live avatars.
Animates a portrait from a reference video: an actor's facial expressions and head turns are transferred to the photo. Runs fast even on a weak GPU.
An older lightweight talking-head model: one photo plus audio becomes a video. Runs on weak hardware, but quality is noticeably below newer models.
The classic lip-to-audio sync model, still popular in hobbyist setups. Lip movements are accurate but the face looks blurry; the license is non-commercial.
A foundation model for market candlestick data: trained on data from more than 45 exchanges, it forecasts prices and volumes. The largest version, large, is not open.
A reasoning model for financial tasks based on Qwen2.5-7B: calculations, report analysis, regulatory questions. Trained on Chinese and English data.
One of the first open model families for financial text: reading statements, news and questions about numbers. Not investment advice: decisions are made by a specialist.
Finance-tuned versions of Llama 2: reading industry texts, reports and questions about terminology. Not investment advice: decisions are made by a specialist.
A classic model that determines the tone of financial news: positive, negative or neutral. English only, runs fast on a CPU.
Ai2's family of models for Sentinel-1, Sentinel-2 and Landsat imagery, with ready-made fine-tunes for mangroves, deforestation and ecosystem types. The license excludes the extractive industries.
A multimodal Earth model: understands optical and radar imagery, terrain, vegetation index and land use maps, and can generate a missing data type (for example, a "see-through-clouds" image from radar).
Foundation models for Landsat and Sentinel-2 satellite imagery that account for image time series. Ready-made fine-tunes for floods, burn scars and crop types, plus a separate weather model, WxC.
A compact research model for satellite imagery, trained on both optical (Sentinel-2) and radar (Sentinel-1) data. Narrower in scope and community than Prithvi and TerraMind.
Pretrained models from the Satlas project for Sentinel-2, Landsat and high-resolution aerial imagery. The predecessor of OlmoEarth, still used in TorchGeo.
NVIDIA's fast image model: 4K images in seconds, runs even on a laptop GPU. The Sprint version generates in 1–2 steps.
The model that started open image generation. A huge ecosystem of fine-tunes, styles and plugins; runs even on a home PC. The popular SDXL-Lightning and Hyper-SD accelerators were made by ByteDance.
An add-on for image models: sets pose, outlines, depth or floor plan so the result follows the required composition exactly.
A compact 0.6B image model with quality on par with much larger ones. The Sigma version does 4K; suits modest hardware.
The family that started "single-pass" 3D reconstruction from a pair or set of photos without camera calibration. MASt3R added point matching and scale; MUSt3R and BLASt3R added video support.
VAST family: a 3D model from a single photo. TripoSR runs in under a second, TripoSG gives cleaner geometry, TripoSplat builds a scene from Gaussian points.
Early open OpenAI models that create a 3D object from text or an image in seconds. Quality is basic, but they are fast and easy to run.
The most widely used open toolkit for face detection and recognition. Many identity-preserving image generators are built on it. The pretrained weights are non-commercial.
Transformer-based face recognition from ByteDance, one of the most accurate open models on benchmarks. Weights are published in ONNX format but are non-commercial.
One of the first adapters that transfer a face from a photo into a generated image. The SD 1.5 versions run on low-end cards, but the weights are non-commercial.
Generates intermediate frames: turns 24–30 fps into 60 fps and more and makes smooth slow motion. Versions 4.24+ smooth out video from generative models well.
A module that brings Stable Diffusion image models to life, turning them into short animations. One of the first open video technologies.
A compact (250M) model for searching document pages as images. According to the authors, it matches models 10 times larger and runs without a GPU.
Searches PDFs and scans as images: pages do not need to be OCR'd first, the model finds the right one for a question directly, including tables and charts. Trained on English.
A small robot control model that runs on a regular laptop. Trained on open data from the LeRobot community, suited to low-cost robot arms.
A lightweight try-on model that runs on a regular GPU. There is a mask-free version and CatV2TON, which also tries clothes on in video.
One of the first weather neural networks, published in Nature and added to ECMWF charts. The weights are open for research only; commercial use is prohibited.
Protein language models: they understand amino acid sequences, predict structure (ESMFold2) and help with protein design. Since 2026 all open versions are under MIT.
An open driver assistance system: a neural network keeps the lane and controls speed from a camera, plus a driver attention monitoring model. The models live right in the repository and are updated constantly.
The bottleneck here is speed. A language model on a CPU answers in seconds rather than instantly, so live customer chat is usually not the place for it. On overnight jobs the difference is invisible. One more point: you need roughly as much RAM as the model file weighs, plus headroom for context.
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.
