AI Open Weight Models

Canonical version: AI Open Weight Models.

AI models whose trained weights are publicly released, allowing anyone to download, run, fine-tune, and deploy them. "Open weight" is more precise than "open source" because most releases don't include training code or data.

Key families: Llama (Meta), Mistral/Mixtral (Mistral AI), Gemma (Google DeepMind), Qwen (Alibaba), Phi (Microsoft), DeepSeek (see DeepSeek v4 — currently the largest open-weights model at 1.6T total / 49B active params), GLM-5.2 (Zhipu AI (Z.ai), 1M context, ~753B/40B; leading open-weight on the AA Intelligence Index at its June 2026 release), Kimi K2 (Moonshot AI, see Kimi K2.6 and Kimi K2.5), Granite (IBM — dense, enterprise-targeted, Apache 2.0).

Significance: democratizes AI access, enables local/private deployment, reduces vendor lock-in, enables community fine-tuning. Tension with safety: open weights can't be un-released, making misuse harder to prevent.

Open-weight model notes

Auto-populated from notes carrying ai/open_weight. Tag a note rather than editing this list to add a model.

• 2026-04-21 Kimi K2.6, Qwen, and Gemma 4 - Local AI Is Catching Up
• 2026-08-01 Kimi K3, Qwen 3.8, and a Week That Reset the Open-Weight Frontier
• Baichuan
• Bonsai 27B
• Georgi Gerganov Machine Learning (GGML)
• GPT-Generated Unified Format (GGUF)
• Granite
• Hermes
• Kimi
• Kimi K2.5
• llama.cpp
• Mistral.rs
• Moonshot AI
• Nous Research
• Safetensors
• Stability AI
• Unsloth
• Zhipu AI (Z.ai)

References

Related

• Large Language Models (LLMs)
• Knowledge Distillation
• Low Rank Adapter (LoRA)
• Atropos
• Tinker
• Axolotl

About Sébastien

Ready to get to the next level?