Open-source AI models are language and other models whose weights are released publicly for anyone to download, fine-tune, and run. This hub tracks releases, licensing, and capabilities from projects like Llama, Mistral, DeepSeek, and Qwen, and what they mean for developers and companies who want control over their own AI infrastructure.
Xiaomi released the code and checkpoints for Xiaomi-Robotics-1, a vision-language-action model it says tops several robot-manipulation benchmarks, under an Apache 2.0 license.
Read more →Mistral AI has open-sourced Shieldstral, a 3-billion-parameter model that moderates text and images against custom, plain-language policies without retraining.
Read more →Thinking Machines Lab's second open-weight model needs a quarter of its predecessor's active parameters yet edges past it on reasoning and coding benchmarks.
Read more →DeepSeek's official V4-Flash-0731 release outperforms its larger V4-Pro model on coding and agent benchmarks, despite activating far fewer parameters per token.
Read more →LG AI Research released K-EXAONE 2.0, a 750-billion-parameter open-weight model under the Apache 2.0 license, aiming to rival Chinese and US frontier AI labs.
Read more →Ant Group's Ling-3.0-Flash matches models several times its size on reasoning and coding benchmarks using just 5.1 billion active parameters, the company says.
Read more →Beijing-based Moonshot AI has published the full open weights for Kimi K3, a 2.8-trillion-parameter model it calls the largest open model released to date.
Read more →
Poolside is a startup that builds AI models trained specifically to write and fix code, then gives the weights away for free. Here's who is behind it and how the models work.
Read more →Poolside's new 118-billion-parameter Laguna S 2.1 matches or beats far larger rivals on agentic coding benchmarks and ships under an open license.
Read more →
FLUX is Black Forest Labs' family of AI image and video models, usable through a free browser Playground, a paid API, or open-weight downloads. Here's what it is and how to try it.
Read more →
Mistral AI is the Paris-based lab racing to build Europe's answer to OpenAI and Anthropic — known for releasing open-weight models and for its 'sovereign AI' pitch to European governments.
Read more →
Kimi K3 is a giant open-weight AI model from China's Moonshot AI. Here's what it is, who built it, and why its release rattled Nvidia and chip investors.
Read more →China's Moonshot AI released Kimi K3, a 2.8-trillion-parameter open model with a 1-million-token context window — a system the company calls the world's first open 3-trillion-class AI model.
Read more →PrismML has released Bonsai 27B, a 27-billion-parameter open-weight model compressed to under 4GB, small enough to run natively on a phone while keeping most of its reasoning ability.
Read more →Mira Murati's startup released Inkling, a 975-billion-parameter open-weight model built for developers to fine-tune and adapt, rather than to top leaderboards.
Read more →
DeepSeek is the Chinese AI lab whose cheaply trained, open-weight models triggered Nvidia's biggest one-day stock drop and reshaped how the industry talks about AI costs.
Read more →Meta now builds two separate AI model families: the open-weight Llama and the new proprietary Muse. Here's how they differ, and how to try either one.
Read more →GLM is the open-weight AI model family from Chinese lab Zhipu AI (Z.ai) — free to download and self-host, and competitive with top US models at a fraction of the API cost.
Read more →Z.ai has released GLM-5.2, a 744-billion-parameter open-weight model that matches GPT-5.5 on key benchmarks and trails Claude Opus 4.8 narrowly — trained entirely on Huawei silicon with no Nvidia hardware.
Read more →Chinese tech company Meituan has disclosed that "Owl Alpha," an anonymous model that topped OpenRouter rankings for two months, was its 1.6-trillion-parameter LongCat-2.0 — trained entirely on domestic chips and now free under an MIT license.
Read more →Mistral's new open-source model for Lean 4 fully saturates the miniF2F theorem benchmark, solves 87% of graduate-level math tests, and found five previously unknown bugs in production code — at roughly $4 per proof versus $300 for competing systems.
Read more →