Text2026
Ricursive Intelligence · USA
A dense (not MoE) model with 65B parameters, pretrained on 12 trillion tokens, with a context of up to 1 million tokens. It is built for programming and agents: by the authors' own test it scores 80.52 on Terminal-Bench 2.1 and beats much larger American open models (Reflection Beam 501B with 80.1, Meta Muse Spark Open 816B with 76.4, NVIDIA Nemotron 3 Ultra 550B with 56.4). The authors started from the Olmo 3 recipe and claim a 16 to 55 times gain in training efficiency on code. It was designed for NVIDIA graphics cards. BF16 and NVFP4 are published for the instruction-tuned version, plus a BF16 base. This is a preview release, and the figures are the authors' own, not independent.
- an agent for programming in the terminal and a repository
- long tasks with a context of hundreds of thousands of tokens
- your own tuning on top of an open base model
- Sizes
- 65B (dense)
- Hardware
- from: Cluster
Text2026
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
Text2026
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
Code2026
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
Text2026
Reflection AI · USA
An open model from the American company Reflection with 501B parameters (23B active, a 1 million token context), text only, mainly English. Stated as economical in reasoning: according to the authors it needs about three times fewer tokens for thinking at comparable quality. Strong in programming, tool calling and science. Weights in BF16, FP8 and NVFP4. The benchmark figures are the authors' own.
- an agent for solving tasks in a repository and a terminal
- long tasks with tools and search (via MCP)
- solving scientific and mathematical problems with reasoning
- Sizes
- 501B-A23B
- Hardware
- from: Cluster
Text2026
Shanghai Artificial Intelligence Laboratory · China
An agent model for long tasks: searching and gathering evidence, research, development, preparing reports and presentations, analyzing vulnerabilities in an authorized environment. A preview based on GLM-5.2, 256K-token context, text only. An FP8 version exists.
- deep search and research on a topic with a verifiable result
- developing programs and preparing reports and presentations
- checking and fixing vulnerabilities in an authorized environment
- Sizes
- 744B (MoE, based on GLM-5.2)
- Hardware
- from: Cluster
Computer-use agents2026
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
Text2026
Aleph Alpha · Germany
A European open reasoning model for German and English. A 78B mixture-of-experts with 3.5B active per token, a 262K-token context (tested up to a million) and tool calling.
- answers over long documents and RAG
- an agent that calls tools
- extracting structured data from documents
- Sizes
- 78B-A3.5B
- Hardware
- from: Cluster
Text2024–2026
MBZUAI, Institute of Foundation Models (IFM, LLM360 project) · UAE
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.
- Reasoning, maths and technical questions
- Analysing long documents
- Agents and writing code
- Sizes
- 0.9B – 375B-A23B
- Hardware
- from: Laptop
Math and reasoning2026
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
Math and reasoningOllama2025–2026
Open Thoughts (Stanford, Berkeley and other universities) · USA
Fully open reasoning models: both weights and training data are published. Newer OpenThinkerAgent versions can carry out multi-step tasks.
- Calculations and formula checks
- Complex analytics with step-by-step breakdowns
- Checking the logic of internal policies
- Sizes
- 1.5B – 32B
- Hardware
- from: Laptop
TextGGUF2025–2026
ServiceNow · USA
ServiceNow 15B models with step-by-step reasoning that fit on a single GPU. From version 1.5 they also understand images and are good at calling tools.
- A reasoning assistant for internal services
- Tool calling and enterprise agents
- Analysing screenshots and documents with images
- Sizes
- 5B – 15B
- Hardware
- from: Laptop
CodeOllama2025–2026
Essential AI · USA
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.
- Writing and fixing code
- A developer agent on a single GPU
- Solving technical and scientific problems
- Sizes
- 8B
- Hardware
- from: Laptop
Math and reasoningGGUF2025–2026
Princeton University · USA
Open models for formal proofs in Lean 4 from Princeton. The new Goedel-Code-Prover proves program correctness.
- Formal verification of mathematical workings
- Verifying code correctness
- Training
- Sizes
- 7B – 32B
- Hardware
- from: Laptop
Math and reasoningGGUF2026
LM Provers (CMU, Hugging Face, ETH Zurich, Project Numina) · USA, Switzerland, France
A small 4B model on Qwen3 that writes mathematical proofs in plain language almost at the level of large models. Runs on a laptop.
- Checking the logic of reasoning and workings
- Step-by-step explanations of solutions
- Training and olympiad preparation
- Sizes
- 4B
- Hardware
- from: Laptop
TextOllama2025
Deep Cogito · USA
Fine-tuned Llama, Qwen and DeepSeek models with a hybrid mode: answer immediately or reason first. The 671B v2.1 flagship spends noticeably fewer tokens on reasoning than DeepSeek R1.
- A chat assistant with a reasoning mode
- Writing code and calling tools
- Answering complex questions about documents
- Sizes
- 3B – 671B
- Hardware
- from: Laptop
Math and reasoningGGUF2024–2025
DeepSeek · China
DeepSeek's maths models. The first 7B version introduced the GRPO training method; the 685B V2 writes and checks its own olympiad-level proofs.
- Calculations and formula checks
- Checking mathematical workings in reports
- Working through problems step by step
- Sizes
- 7B – 685B
- Hardware
- from: Laptop
TextOllama2023–2025
Nous Research · USA
Nous Research fine-tunes on top of Llama, Mistral, Qwen and Seed-OSS. Valued for precise instruction following, function calling and strict JSON output; they refuse less often than the originals; Hermes 4 has a reasoning mode.
- Agents that call functions and APIs
- Data extraction in strict JSON format
- Assistant with flexible role and tone settings
- Sizes
- 3B – 405B
- Hardware
- from: Laptop
Moderation and safetyOllama2025
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
Math and reasoningGGUF2025
Moonshot AI and Project Numina · China, France
Models for formal proofs in Lean 4 from Moonshot AI (Kimi) and Numina. Small versions from 0.6B run on a laptop.
- Formal verification of mathematical workings
- Translating a problem from plain language into Lean
- Training and olympiad preparation
- Sizes
- 0.6B – 72B
- Hardware
- from: Laptop
TextRU2024–2025
Lomonosov Moscow State University Research Computing Center, LAIR lab (RefalMachine) · Russia
Qwen models adapted for Russian: a new tokenizer plus further training on Russian texts. As a result, Russian text is generated up to twice as fast as with the original model of the same size.
- Russian-language assistant on your own server
- Answers based on company documents (RAG) in Russian
- Analysis and summaries of long Russian texts
- Sizes
- 1.5B – 32B
- Hardware
- from: Laptop
Math and reasoningOllama2025
Agentica (Berkeley, Sky Computing Lab) and Together AI · USA
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.
- Solving maths problems with step-by-step working
- Generating and checking code
- An agent for fixing bugs in a repository
- Sizes
- 1.5B – 32B
- Hardware
- from: Laptop
MedicineGGUF2025
Intelligent Internet · UK
Reasoning medical models on Qwen3, designed to run on an ordinary computer. Does not replace a doctor; decisions are made by a specialist.
- Reference answers to staff with the reasoning shown
- Draft discharge summaries for a doctor to review
- Searching medical literature
- Sizes
- 7B – 32B
- Hardware
- from: Laptop
Math and reasoning2025
NVIDIA · USA
NVIDIA models for maths and reasoning based on Qwen. AceReason was fine-tuned with reinforcement learning first on maths, then on code.
- Calculations and formula checks
- Complex analytics with step-by-step breakdowns
- Working through programming problems
- Sizes
- 1.5B – 72B
- Hardware
- from: Laptop
Text to SQLGGUF2025
IDEA Research · China
A text-to-SQL model trained with reinforcement learning: it works through the schema and the conditions step by step before producing a query.
- Database queries for questions with several conditions
- Reviewing and fixing other people SQL queries
- An analyst helper inside a BI system
- Sizes
- 3B – 14B
- Hardware
- from: Laptop
Finance2025
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
Math and reasoningGGUF2024–2025
DeepSeek · China
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.
- Formal verification of mathematical workings
- Verifying algorithm correctness
- Training and olympiad preparation
- Sizes
- 7B – 671B
- Hardware
- from: Laptop
Finance2025
Shanghai University of Finance and Economics (SUFE) · China
A reasoning model for financial tasks based on Qwen2.5-7B: calculations, report analysis, regulatory questions. Trained on Chinese and English data.
- Financial calculations with step-by-step explanations
- Answering questions about financial statements
- Analyzing tables of financial data
- Sizes
- 7B
- Hardware
- from: Laptop
Math and reasoningGGUF2025
Qihoo 360 · China
Reasoning models from Qihoo 360: a standard Qwen2.5 was fine-tuned for long reasoning using an open recipe; data and code are published.
- Calculations and formula checks
- Working through problems step by step
- A base for your own reasoning fine-tuning
- Sizes
- 7B – 32B
- Hardware
- from: Laptop
Math and reasoningOllama2024–2025
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
Math and reasoningGGUF2025
Stanford University · USA
A reasoning model trained on just a thousand problems. It can be told to think longer to answer a hard question more accurately.
- Calculations and formula checks
- Working through complex problems step by step
- Training
- Sizes
- 1.5B – 32B
- Hardware
- from: Laptop
Math and reasoningGGUF2025
NovaSky (Sky Computing Lab, Berkeley) · USA
A Berkeley reasoning model trained for under 450 dollars. It showed that o1-preview-level reasoning can be reproduced with modest resources.
- Calculations and formula checks
- Working through problems step by step
- A base for your own reasoning fine-tuning
- Sizes
- 7B – 32B
- Hardware
- from: Laptop
TextOllama2023–2025
Ai2 · USA
Ai2 fine-tunes of Llama with a fully open recipe: data, code and all intermediate stages. Tulu 3 405B is one of the largest openly fine-tuned models; OLMo chat versions use the same recipe.
- Employee assistant on your own server
- Math and precise instruction following
- Reference recipe for your own fine-tuning
- Sizes
- 7B – 405B
- Hardware
- from: Laptop
Math and reasoningOllama2024–2025
Qwen (Alibaba) · China
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.
- Calculations and formula checks
- Checking calculations in estimates and reports
- Working through problems step by step
- Sizes
- 1.5B – 72B
- Hardware
- from: Laptop
TextOllamaNot maintained2023–2024
WizardLM (Microsoft and Peking University) · USA / China
Fine-tunes of Llama, Mistral and StarCoder using Evol-Instruct, which automatically makes instructions more complex. WizardLM-2 was released in April 2024 and removed almost immediately, so only the 2023 versions are relevant.
- Complex multi-step instructions
- Help for developers
- Solving math problems
- Sizes
- 7B – 70B
- Hardware
- from: Laptop
TextOllamaNot maintained2023
Microsoft Research · USA
Microsoft research models based on Llama 2, trained to choose a reasoning approach for each task. The orca-mini model in Ollama is a different project by independent developer Pankaj Mathur.
- Research on reasoning methods
- Comparison with modern small models
- Training specialists
- Sizes
- 7B – 13B
- Hardware
- from: Laptop
