The teams building AI in the open.
Open weights are not always open source.
Downloadable weights let you run and adapt a model. A fully open release also exposes the data, code, checkpoints and training decisions needed to inspect or reproduce it. These profiles label that difference.
7 labs to know
Latest tracked releasesMoonshot AI
Moonshot AI builds the Kimi family at the frontier of long-context reasoning, coding and autonomous agent work.
Z.ai
Z.ai develops the GLM family, with recent releases focused on long-horizon coding, autonomous engineering and efficient sparse attention.
MiniMax
MiniMax builds multimodal frontier models for coding, computer use and agentic work, alongside video, speech and music systems.
Mistral AI
Mistral builds efficient European open models spanning local deployment, reasoning, multimodality and agentic coding.
DeepSeek
DeepSeek focuses on efficient mixture-of-experts systems, reasoning, coding and agentic tasks, with unusually permissive model releases.
Qwen
Alibaba Cloud’s Qwen team builds multilingual, multimodal and agent-focused models across a wide range of deployable sizes.
Google DeepMind
The Gemma programme turns Google DeepMind research into compact, downloadable models for local development and experimentation.