cleverhack.com

OSS model release timeline ↴

The Urgency of Open Source AI

August 2025 (Updated April 2026)

Recent Major Open Source Model Release Timeline

March 2026

NVIDIA Nemotron 3 Super (March 2026) - US — A 12B active / 120B total parameter hybrid Mamba-Transformer MoE model. First in the Nemotron 3 series to use LatentMoE, MTP layers for native speculative decoding, and NVFP4 pretraining. Achieves up to 2.2x higher inference throughput than GPT-OSS-120B and 7.5x over Qwen3.5-122B, while matching or exceeding both on accuracy. Supports up to 1M token context. Pre-trained, post-trained, and quantized checkpoints, plus training datasets and model recipes, are all released. Released under the NVIDIA Open Model license.
https://research.nvidia.com/labs/nemotron/Nemotron-3-Super/

- Alibaba Cloud (Qwen) Qwen 3.5 Small Model Series (March 2026) - China — This latest small model series Qwen3.5-0.8B, Qwen3.5-2B, Qwen3.5-4B, Qwen3.5-9B brings the Qwen3.5 foundation of native multimodal, improved architecture, scaled RL to edge deployment and lightweight agents — the 9B variant already closes the gap with much larger models. Base models are also released to support research and experimentation. Supports 262K native context (extensible to 1M+), 201 languages, and thinking mode with chain-of-thought reasoning. Released under an Apache 2.0 license.
🔗: https://huggingface.co/collections/Qwen/qwen35

February 2026

- Alibaba Cloud (Qwen) Qwen3.5-397B-A17B (February 2026) - China — Qwen3.5 is Alibaba Cloud's latest flagship open-weight model, released February 16, 2026, and built for the "agentic AI era." It uses a sparse Mixture-of-Experts (MoE) architecture with 397B total parameters and 17B active parameters per forward pass. The model is natively multimodal (text, images, audio, video) and features visual agentic capabilities for autonomous interaction across mobile and desktop interfaces. Alibaba reports 60% lower inference costs and 8x higher throughput versus its predecessor. Benchmark highlights include 83.6 on LiveCodeBench v6, 91.3 on AIME26, and 76.4 on SWE-bench Verified. Released under an Apache 2.0 license; a closed-source Qwen3.5-Plus variant with a 1M-token context window is available via Alibaba Cloud.
🔗: https://huggingface.co/Qwen/Qwen3.5-397B-A17B

- Zhipu AI (Z.ai) GLM-5 (February 2026) - China — GLM-5 is the fifth-generation large language model from Zhipu AI (now Z.ai), a Tsinghua University spinoff that completed a landmark Hong Kong IPO in January 2026. It is a 745B-parameter sparse Mixture-of-Experts (MoE) model with approximately 40-44B active parameters per token across 256 experts, 8 activated per token. GLM-5 supports a 200K-token context window using DeepSeek Sparse Attention and was trained entirely on Huawei Ascend chips using the MindSpore framework — zero dependency on US-manufactured hardware. The model targets coding, advanced reasoning, and agentic intelligence, scoring 77.8% on SWE-bench Verified and ranking as the top open-source model on Vending Bench 2 and BrowseComp. GLM-5 also achieved a record-low hallucination rate on the Artificial Analysis Intelligence Index v4.0. Released under an MIT license.
🔗: https://huggingface.co/zai-org/GLM-5

- Cohere Labs Tiny Aya (February 2026) - Canada / US — Described as the most capable multilingual open-weight model at the 3B parameter class, outperforming Qwen3-4B, Gemma 3 4B, and Ministral 3B. A classic decoder-style transformer optimized for multilingual understanding across dozens of languages. Released under a non-commercial CC-BY-NC-4.0 license.
🔗: https://huggingface.co/CohereLabs/tiny-aya-base

- MiniMax M2.5 (February 2026) - China — MiniMax M2.5 is the latest flagship model from MiniMax, a Chinese AI company that IPO'd on the Hong Kong Stock Exchange in early 2026. Described as "the world's first production-level model designed natively for Agent scenarios," M2.5 is a Mixture-of-Experts model optimized for coding and agentic workflows. It achieves 80.2% on SWE-bench Verified, edging out GPT-5.2 (80.0%) and Gemini 3 Pro (78%), while approaching Claude Opus 4.6 (80.8%). The model also scored 76.3% on BrowseComp and demonstrates SOTA performance in office productivity tasks. M2.5-Lightning variant delivers 100 tokens per second at $0.3/M input and $2.4/M output tokens. MiniMax has begun internal testing of its overseas Agent product with M2.5 integrated. Released under a Modified-MIT license.
🔗: https://www.minimax.io/news/minimax-m25

January 2026

- Arcee AI Trinity Large (January 2026) - US — Trinity-Large-Base is a pretrained foundation model from Arcee AI's Trinity Large training run. It is a 398B-parameter sparse Mixture-of-Experts (MoE) model with approximately 13B active parameters per token. The checkpoint was captured after 17 trillion tokens of pretraining, including mid-training learning-rate anneals and context extension, but prior to any instruction tuning or reinforcement learning. This checkpoint represents the completed pretraining phase and serves as a foundation for research and downstream fine-tuning. Released under an Apache 2.0 license.
🔗: https://www.arcee.ai/blog/trinity-large

- Ai2 SERA (January 2026) - US — SERA-32B is the first model in Ai2's Open Coding Agents series. It is a state-of-the-art open-source coding agent that achieves 49.5% on SWE-bench Verified, matching the performance of frontier open models like Devstral-Small-2 (24B) and larger models like GLM-4.5-Air (110B). SERA-32B was trained using Soft Verified Generation (SVG), a simple and efficient method that is 26x cheaper than reinforcement learning and 57x cheaper than previous synthetic data methods to reach equivalent performance. The total cost for data generation and training is approximately $2,000 (40 GPU-days). Released under an Apache 2.0 license.
🔗: https://huggingface.co/allenai/SERA-32B

- Moonshot AI Kimi-K2.5 (January 2026) - China — Kimi K2.5 is an open-source, native multimodal agentic model built through continual pretraining on approximately 15 trillion mixed visual and text tokens atop Kimi-K2-Base. It seamlessly integrates vision and language understanding with advanced agentic capabilities, instant and thinking modes, as well as conversational and agentic paradigms. Released under a Modified MIT license.
🔗: https://huggingface.co/moonshotai/Kimi-K2.5

- Liquid AI LFM2.5 (January 2026) - US — LFM2.5 is a family of 1B parameter models enabling access to private, fast, and always-on intelligence on any device. Our new Text models offer uncompromised quality for high-performance on-device workflows. The Audio model is 8x faster than its predecessor, running natively on constrained hardware like vehicles, mobiles, and IoT devices. Finally, our VLM boosts multi-image, multilingual vision understanding and instruction following for on-edge multimodal use. Released under the LFM Open License v1.0.
🔗: https://www.liquid.ai/blog/introducing-lfm2-5-the-next-generation-of-on-device-ai

- Z.ai GLM-4.7-Flash (January 2026) - China — Your local coding and agentic assistant. Setting a new standard for the 30B class, GLM-4.7-Flash balances high performance with efficiency, making it the perfect lightweight deployment option. Beyond coding, it is also recommended for creative writing, translation, long-context tasks, and roleplay. Released under an MIT license.
🔗: https://huggingface.co/zai-org/GLM-4.7-Flash

December 2025

- MiniMax MiniMax-M2.1 (December 2025) - China — MiniMax-M2.1 is a new open-source AI model with 10 billion activated parameters (230 billion total) that democratizes high-performance agentic capabilities, scoring 74.0 on SWE-bench Verified and 91.5 on VIBE-Web benchmarks. It excels in multi-language programming (Rust, Java, Go, C++, TypeScript, etc.), UI development, and complex real-world office workflows while offering full transparency and accessibility through both HuggingFace weights and API access. Released under a Modified-MIT license.
🔗: https://huggingface.co/MiniMaxAI/MiniMax-M2.1

- Z.ai GLM-4.7 (December 2025) - China — Optimized for AI coding assistance, this updated model shows major improvements over GLM-4.6 across coding tasks (including 5.8% gain on SWE-bench and 12.9% on multilingual coding), UI/webpage generation, tool usage, and complex reasoning. It also delivers better performance in chat, creative writing, and role-play scenarios. Released under an MIT license.
🔗: https://huggingface.co/zai-org/GLM-4.7

- NVIDIA Nemotron 3 (December 2025) - US — The Nemotron 3 family consists of three models: Nano, Super, and Ultra. These models deliver strong agentic, reasoning, and conversational capabilities. Features include a hybrid Mamba-Transformer MoE architecture (with Latent MoE and Multi-Token Prediction layers in Super and Ultra variants), trained using NVFP4 and multi-environment reinforcement learning for superior accuracy across diverse tasks. These models support up to 1M token context windows and offer granular reasoning budget control at inference time. Released under the NVIDIA Open Model license.
🔗: https://research.nvidia.com/labs/nemotron/Nemotron-3/?ncid=ref-inor-399942

- Essential AI rnj-1-instruct (December 2025) - US — Rnj-1 is a family of 8B parameter open-weight, dense models trained from scratch by Essential AI, optimized for code and STEM with capabilities on par with SOTA open-weight models. These models perform well across a range of programming languages and boast strong agentic capabilities (e.g., inside agentic frameworks like mini-SWE-agent), while also excelling at tool-calling. They additionally exhibit strong capabilities in math and science. Released under an Apache 2.0 license.
🔗: https://huggingface.co/EssentialAI/rnj-1-instruct

- MBZUAI K2 (K2-V2) (December 2025) - UAE — A 360-open LLM built from scratch as a superior base for reasoning adaptation, while still excelling at core LLM capabilities like conversation, knowledge retrieval, and long-context understanding. With a 70B dense transformer architecture engineered as a reasoning-enhanced base model and native 512K context (extendable via RoPE scaling). Instead of releasing only weights, we’re sharing the full training story — dataset recipes, mid-training checkpoints, logs, code, and evaluation tools. Released under an Apache 2.0 license.
🔗: https://huggingface.co/LLM360/K2-V2

- Mistral AI Mistral 3 (December 2025) - France / US — Mistral 3 includes three state-of-the-art small, dense models (14B, 8B, and 3B) and Mistral Large 3 – the most capable model to date – a sparse mixture-of-experts trained with 41B active and 675B total parameters. All models are released under the Apache 2.0 license. The Ministral models represent the best performance-to-cost ratio in their category. At the same time, Mistral Large 3 joins the ranks of frontier instruction-fine-tuned open-source models.
🔗: https://mistral.ai/news/mistral-3

- Arcee AI Trinity (December 2025) - US — Trinity Mini is a compact MoE model trained end-to-end in the U.S., offering open weights, strong reasoning, and full control for developers. Part of the new Trinity family, a series of open-weight foundation models for enterprise and tinkerers alike. Released under an Apache 2.0 license.
🔗: https://huggingface.co/arcee-ai/Trinity-Mini

- DeepSeek-V3.2 (December 2025) - China — This model harmonizes high computational efficiency with superior reasoning and agent performance. Featuring three key technical breakthroughs: DeepSeek Sparse Attention (DSA), Scalable Reinforcement Learning Framework, and a Large-Scale Agentic Task Synthesis Pipeline. Performs comparably to GPT-5. The high-compute variant, DeepSeek-V3.2-Speciale, surpasses GPT-5 and exhibits reasoning proficiency on par with Gemini-3.0-Pro. Achieved Gold-medal performance in the 2025 International Mathematical Olympiad (IMO) and International Olympiad in Informatics (IOI). Released under a MIT license.
🔗: https://huggingface.co/deepseek-ai/DeepSeek-V3.2

November 2025

- Prime Intellect INTELLECT-3 (November 2025) - US — INTELLECT-3 is a 106B (A12B) parameter Mixture-of-Experts reasoning model post-trained from GLM-4.5-Air-Base using supervised fine-tuning (SFT) followed by large-scale reinforcement learning (RL). Training was performed with prime-rl using environments built with the verifiers library. All training and evaluation environments are available on the Environments Hub. The model, training frameworks, and environments are open-sourced under fully-permissive licenses (MIT and Apache 2.0).
🔗: https://huggingface.co/PrimeIntellect/INTELLECT-3

- DeepSeekMath-V2 (November 2025) - China — Demonstrates strong theorem-proving capabilities, achieving gold-level scores on IMO 2025 and CMO 2024 and a near-perfect 118/120 on Putnam 2024 with scaled test-time compute. Released under an Apache 2.0 license.
🔗: https://huggingface.co/deepseek-ai/DeepSeek-Math-V2

- Microsoft Fara-7B (November 2025) - US — Microsoft's first agentic small language model for computer use. This experimental model includes robust safety measures to aid responsible deployment. With only 7 billion parameters, Fara-7B is an ultra-compact Computer Use Agent (CUA) that achieves state-of-the-art performance within its size class and is competitive with larger, more resource-intensive agentic systems. Released under a MIT license.
🔗: https://huggingface.co/microsoft/Fara-7B

- Ai2 Olmo 3-Think (32B) (November 2025) - US — The best fully open 32B-scale thinking model that for the first time lets you inspect intermediate reasoning traces and trace those behaviors back to the data and training decisions that produced them. Olmo 3 is a family of compact, dense models at 7 billion and 32 billion parameters that can run on everything from laptops to research clusters. Released under an Apache 2.0 license.
🔗: https://huggingface.co/allenai/Olmo-3-1125-32B

- Motif Technologies Motif-2-12.7B-Instruct (November 2025) - South Korea — Designed for scalable language understanding and robust instruction generalization under constrained compute budgets, the Motif-2-12.7B model family demonstrates competitive performance across diverse benchmarks, showing that thoughtful architectural scaling and optimized training design can rival the capabilities of much larger models. Released under an Apache 2.0 license.
🔗: https://huggingface.co/Motif-Technologies/Motif-2-12.7B-Instruct

- Deep Cogito Cogito v2.1 (November 2025) - US — Open weight model forked off the open-licensed Deepseek base model from November 2024. Weights released under a MIT license.
🔗: https://www.deepcogito.com/research/cogito-v2-1

- Meta FAIR Omnilingual Automatic Speech Recognition (ASR) (November 2025) - US - A suite of models providing automatic speech recognition capabilities for more than 1,600 languages, achieving state-of-the-art quality at an unprecedented scale. Designed as a community-driven framework. Released under Apache 2.0 license while the data is provided under the CC-BY license.
🔗: https://ai.meta.com/blog/omnilingual-asr-advancing-automatic-speech-recognition/

- Moonshot AI Kimi K2 Thinking (November 2025) - China — Built as a thinking agent, it reasons step by step while using tools, achieving state-of-the-art performance on Humanity's Last Exam (HLE), BrowseComp, and other benchmarks, with major gains in reasoning, agentic search, coding, writing, and general capabilities. It can execute up to 200 – 300 sequential tool calls without human interference. Released under a Modified-MIT license.
🔗: https://huggingface.co/moonshotai/Kimi-K2-Thinking

October 2025

- Facebook MobileLLM-P1 (October 2025) - US — MobileLLM-P1 or Pro, a 1B foundational language model in the MobileLLM series, designed to deliver high-quality, efficient on-device inference across a wide range of general language modeling tasks. Released under a FAIR Noncommercial Research License.
🔗: https://huggingface.co/facebook/MobileLLM-Pro

- ByteDance Ouro (October 2025) - China — Ouro-1.4B is a 1.4 billion parameter Looped Language Model (LoopLM) that achieves exceptional parameter efficiency through iterative shared-weight computation. Released under an Apache 2.0 license.
🔗: https://huggingface.co/ByteDance/Ouro-1.4B

- MiniMax MiniMax-M2 (October 2025) - China — It's a compact, fast, and cost-effective MoE model (230 billion total parameters with 10 billion active parameters) built for elite performance in coding and agentic tasks, all while maintaining powerful general intelligence. At only 8% of the price of Claude Sonnet and twice the speed. Released under a MIT license.
🔗: https://huggingface.co/MiniMaxAI/MiniMax-M2

- ServiceNow Apriel-1.5-15b-Thinker (October 2025) - US — A multimodal reasoning model in ServiceNow’s Apriel SLM series which achieves competitive performance against models 10 times it's size. Released under a MIT license.
🔗: https://huggingface.co/ServiceNow-AI/Apriel-1.5-15b-Thinker

- IBM Granite 4.0 (October 2025) - US — A new era for IBM’s family of enterprise-ready large language models, leveraging novel architectural advancements to enable small, efficient language models that provide competitive performance at reduced costs and latency. Multilingual and developed with a particular emphasis on essential tasks for agentic workflows. Released under an Apache 2.0 license.
🔗: https://huggingface.co/collections/ibm-granite/granite-40-language-models-6811a18b820ef362d9e5a82c

September 2025

- Z.ai GLM-4.6 (September 2025) - China — This latest model features a longer context window, superior coding performance, advanced reasoning, more capable agents, and refined writing versus GLM-4.5. Released under a MIT license.
🔗: https://huggingface.co/zai-org/GLM-4.6

- Alibaba Qwen3-Omni (September 2025) - China — A native end-to-end multilingual omni-modal SOTA foundation model family. Processes text, images, audio, and video, and delivers real-time streaming responses in both text and natural speech. Supports 119 text languages, 19 speech input languages, and 10 speech output languages. Released under an Apache 2.0 license.
🔗: https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Instruct

- MBZUAI & G42 K2-Think (September 2025) - UAE / US — K2-Think is a 32 billion parameter open-weights general reasoning model with strong performance in competitive mathematical problem solving. This model is the result of a joint effort between the Institute of Foundation Models at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) and G42. Released under an Apache 2.0 License.
🔗: https://huggingface.co/LLM360/K2-Think

- Swiss AI Initative Apertus (September 2025) - Switzerland — Apertus is a fully-open 70B and 8B parameter language model supporting over 1000 global languages (including Swiss German and Romansh), long context, using only fully compliant and open training data, all while achieving comparable performance to models trained behind closed doors. Released under an Apache 2.0 License.
🔗: https://huggingface.co/swiss-ai/Apertus-70B-2509

August 2025

- Meituan LongCat-Flash-Chat (August 2025) - China — A non-thinking foundation model that delivers highly competitive performance, with exceptional strengths in agentic tasks. Released under a MIT license.
🔗: https://huggingface.co/meituan-longcat/LongCat-Flash-Chat

- Nous Research Hermes 4 (August 2025) - US — Frontier, steerable, open, hybrid reasoning 405b and 70b model variants based on Meta's Llama-3.1. Strong performance in math, coding, STEM, and creativity. Released under a Llama 3 Community License.
🔗: https://hermes4.nousresearch.com/

- Cohere Labs Command A Reasoning (August 2025) - Canada / US — An open weights research release of a 111 billion parameter model optimized for tool use, agentic, and multilingual use cases with reasoning capabilities. Released under a CC-BY-NC-4.0 license.
🔗: https://huggingface.co/CohereLabs/command-a-reasoning-08-2025

- DeepSeek DeepSeek-V3.1 (August 2025) - China — Combines earlier V3 and R1 models into a hybrid thinking/non-thinking reasoning model with improved tool use and agentic capabilities. Released under a MIT license.
🔗: https://huggingface.co/deepseek-ai/DeepSeek-V3.1-Base

- ByteDance Seed-OSS (August 2025) - China — A series of open-source LLMs designed for long-context, reasoning, agent, and general capabilities with developer-friendly features. Seed-OSS achieves excellent performance on several popular open benchmarks. Released under an Apache 2.0 license.
🔗: https://huggingface.co/ByteDance-Seed/Seed-OSS-36B-Instruct

- NVIDIA Nemotron-Nano-9B-v2 (August 2025) - US — New small, open model for enterprises with toggle on/off reasoning and open weights, open datasets, and training techniques. Released under the NVIDIA Open Model license.
🔗: https://huggingface.co/blog/nvidia/supercharge-ai-reasoning-with-nemotron-nano-2

- Google Gemma 3 270M (August 2025) - US — A compact, 270-million parameter model designed for task-specific fine-tuning with strong instruction-following and text structuring capabilities already trained in. Released under the Gemma license.
🔗: https://developers.googleblog.com/en/introducing-gemma-3-270m/

- OpenAI gpt-oss-120B and gpt-oss-20B (August 2025) - US — Open-weight AI models designed for powerful reasoning, agentic tasks, and versatile developer use cases. Ability to fine-tune. Released under an Apache 2.0 license.
🔗: https://huggingface.co/openai/gpt-oss-120b

July 2025

- Z.ai GLM-4.5 (July 2025) - China — The GLM-4.5 series models are foundation models designed for intelligent agents. Blog post with benchmarks. Released under a MIT license.
🔗: https://huggingface.co/zai-org/GLM-4.5

- Moonshot AI Kimi-K2-Instruct (July 2025) - China — Kimi K2 is a state-of-the-art mixture-of-experts (MoE) language model. Performs across frontier knowledge, reasoning, and coding tasks while optimized for agentic capabilities. Released with a modified-MIT license.
🔗: https://huggingface.co/moonshotai/Kimi-K2-Instruct

- StepFun AI Step3 (July 2025) - China — Step3 is a cutting-edge multimodal reasoning model—built on a Mixture-of-Experts architecture with 321B total parameters and 38B active. Released with an Apache 2.0 license.
🔗: https://huggingface.co/stepfun-ai/step3

- Alibaba Qwen3-Coder (July 2025) - China — Agentic code model with tool calling, available in multiple sizes. Released under an Apache 2.0 license.
🔗: https://qwenlm.github.io/blog/qwen3/

- Mistral Large 2 (July 2025) - France / US — Dense 123B-parameter multilingual model optimized for single-node inference; released under a research/restricted license that allows research use with commercial deployment requiring separate licensing.
🔗: https://mistral.ai/news/mistral-large-2407

June 2025

- Magistral Small (June 2025) - France / US — Mistral’s open-weight reasoning model focused on transparent logical reasoning. Released under an Apache 2.0 license.
🔗: https://mistral.ai/news/magistral

- MiniMax MiniMax-M1 (June 2025) - China — The world's first open-weight, large-scale hybrid-attention reasoning model. Released under an Apache 2.0 license.
🔗: https://huggingface.co/MiniMaxAI/MiniMax-M1-80k

May 2025

- NVIDIA Llama-3.3-Nemotron-Super-49B-v1 (May 2025) - US — This LLM is a derivative of Meta Llama-3.3-70B-Instruct (AKA the reference model). It is a reasoning model that is post trained for reasoning, human chat preferences, and tasks, such as RAG and tool calling. Released under the NVIDIA Open Model license.
🔗: https://huggingface.co/nvidia/Llama-3_3-Nemotron-Super-49B-v1

- DeepSeek-R1-0528 (May 2025) - China — Chinese reasoning model with improved complex reasoning and reduced hallucinations. Released under a MIT license.
🔗: https://huggingface.co/deepseek-ai/DeepSeek-R1-0528

April 2025

- Meta Llama 4 (April 2025) - US — Introduces a mixture-of-experts (MoE) architecture for efficiency and capability scaling; distributed as open-weight with nuanced source-available terms, making its openness debated.
🔗: https://ai.meta.com/blog/llama-4-multimodal-intelligence/

- Alibaba Qwen3 (April 2025) - China — Hybrid dense and MoE reasoning model family aimed at deep reasoning and multimodal tasks; positioned as open-weight/open-source, refer to individual model licensing.
🔗: https://qwenlm.github.io/blog/qwen3/

- IBM Granite 3.3 (April 2025) - US — Enterprise-focused models with enhanced reasoning and deployment transparency; released under an Apache 2.0 license.
🔗: https://huggingface.co/collections/ibm-granite/granite-33-language-models-67f65d0cca24bcbd1d3a08e3

March 2025

- DeepSeek-V3 (March 2025) - China — DeepSeek-V3 is a strong Mixture-of-Experts (MoE) language model. Released under a MIT license.
🔗: https://huggingface.co/deepseek-ai/DeepSeek-V3

- Mistral Small 3.1 / instruction-tuned variant (March 2025) - France / US — Efficient 24B-parameter instruction-tuned model with strong generative capabilities; released under an Apache 2.0 license.
🔗: https://huggingface.co/mistralai/Mistral-Small-24B-Instruct-2501

- Cohere Labs Command-A (March 2025) - Canada / US — Cohere Labs Command A is an open weights research release of a 111 billion parameter model optimized for enterprises. Released under a CC-BY-NC-4.0 license.
🔗: https://huggingface.co/CohereLabs/c4ai-command-a-03-2025

- Google Gemma 3 (March 2025) - US — Open-weight multilingual and vision-language family (1B–27B sizes) with structured output focus; publicly distributed with usage guardrails. This model can run on a single GPU or TPU. Released under a Gemma license.
🔗: https://blog.google/technology/developers/gemma-3/

January 2025

- DeepSeek-R1 (January 2025) - China — DeepSeek-R1 achieves performance comparable to OpenAI-o1 across math, code, and reasoning tasks. Released under a MIT license.
🔗: https://huggingface.co/deepseek-ai/DeepSeek-R1

Previous Notable Releases

- BLOOM (176B, original 2022 release; community updates ongoing) – Multilingual model from the BigScience project; open-access with a Responsible AI–style license encouraging ethical usage.
🔗: https://bigscience.huggingface.co/blog/bloom

- Stable Diffusion variants (2024–2025) – Community-driven open-weight image generation models (e.g., Stable Diffusion 3 and SDXL Lightning) with publicly released weights and tooling; licensing and availability vary across submodels.
🔗: https://stability.ai/news/stable-diffusion-3

Open Source License Definitions

[MIT] MIT License – OSI-approved permissive open-source license allowing modification, redistribution, and commercial use with attribution.

[Apache] Apache 2.0 – OSI-approved permissive open-source license with patent grant and wide reuse rights.

[CC-BY-NC-4.0] CC-BY-NC-4.0 – A Creative Commons license that allows others to share and adapt a work, but only for non-commercial purposes and with proper attribution to the original creator.

[Open-weight / source-available] Trained weights are publicly shared, but associated code or terms may impose limitations; the line between fully open source and source-available is debated.

[Research/restricted] Limited to research/non-commercial use unless additional commercial licensing is obtained.

[Responsible AI–style] Access coupled with normative guidance on usage; not a standard OSI license.

[Community/open-weight ecosystem] Projects like Stable Diffusion combine public weights with community tooling; specific submodels may vary in terms.

Open Source AI Definitions

[Open Source AI] The Open Source Foundation's version 1.0 definition of Open Source AI.

[Open Weights] A discussion about Open Weights from the Open Source Foundation.

Open Source AI Leaderboards

[LMArena.ai] February 2026: Top 10 Open Models in Text

[Artificial Analysis] LLM Leaderboard - Open Source

[Artificial Analysis] Artificial Analysis Openness Index - A composite measure providing an industry standard to communicate model openness for users and developers

[LMArena.ai] January 2026: Top 10 Open Models in Text

[LMArena.ai] EOY 2025: Top 10 Open Models in Text

[LMArena.ai] November 2025: Top 10 Open Models by Provider (Text)

[LMArena.ai] October 2025: Top 10 Open Models by Provider (Text)

[LMArena.ai] September 2025: Top 10 Open Models by Provider (Text)

[LMArena.ai] August 2025: Top 10 Open Models by Provider (Text)

[Interconnects] Ranking the Chinese Open Model Builders - August 2025

Additional Open Source AI Resources

[Adina Yakefu & Irene Solaiman] One Year Since the “DeepSeek Moment” - January 2026

[Nathan Lambert] Relative Adoption Metric (RAM) - January 2026

[Robert Praas, Pierre-Alexandre Balland, Francisco Ríos] Chinese developers account for over 45% of top open-model public downloads - AI World - December 2025

[Nathan Lambert, Florian Brand] 2025 Open Models Year in Review - December 2025

[Loïck BOURDOIS] Model statistics of the 50 most downloaded entities on Hugging Face - October 2025

[Nathan Lambert, et al.] Twitter thread about major open source model milestones - October 2025

[The Economist] China is quietly upstaging America with its open models - August 2025

[The ATOM Project] A new initiative to reinvigorate AI research in the U.S. by building leading, open models. More info.

[Chinese LLM Community] A Hugging Face community tracking open source Chinese LLM releases by month.

Menu	About The Author	More AI Writing
⇒ HOME ⇒ ABOUT	Joy Larkin is a technologist in Silicon Valley. She likes robots and is excited for Superintelligence. LinkedIn: /in/joylarkin ◦◦◦ Twitter: @joy	The Great Frontier AI Lab Product War Is Just Beginning Frontier AI Lab Tracker