Open Weight Models
Open Weight Models · Detailed References

Twenty models, analyzed as infrastructure.

Every model now has a human-readable OWM reference, a machine-readable Passport and an observed change log. The goal is depth, traceability and deployment usefulness — not the largest list.

Cohere Labs

Command A

Detailed OWM reference with license reality, hardware, runtimes, sovereignty, fit guidance and observed change history.

111B256K (HF config may use 128K)CC BY-NC 4.0 + Cohere Labs AUP
DeepSeek

DeepSeek-R1

Detailed OWM reference with license reality, hardware, runtimes, sovereignty, fit guidance and observed change history.

Large DeepSeek-V3-family MoESee current serving configurationMIT
DeepSeek

DeepSeek-R1-Distill-Qwen-32B

Detailed OWM reference with license reality, hardware, runtimes, sovereignty, fit guidance and observed change history.

33B reported on model hub131,072MIT
DeepSeek

DeepSeek-R1-Distill-Qwen-7B

Detailed OWM reference with license reality, hardware, runtimes, sovereignty, fit guidance and observed change history.

7B classSee exact model cardMIT
Mistral AI / All Hands AI

Devstral Small 1.1

Detailed OWM reference with license reality, hardware, runtimes, sovereignty, fit guidance and observed change history.

24B128KApache 2.0
Google DeepMind

Gemma 3 27B IT

Detailed OWM reference with license reality, hardware, runtimes, sovereignty, fit guidance and observed change history.

27B128KGemma Terms
Z.ai

GLM-4.5

Detailed OWM reference with license reality, hardware, runtimes, sovereignty, fit guidance and observed change history.

355B total / 32B activeSee current model cardMIT
OpenAI

gpt-oss-120b

Detailed OWM reference with license reality, hardware, runtimes, sovereignty, fit guidance and observed change history.

120B classSee current model cardApache 2.0 + usage policy
OpenAI

gpt-oss-20b

Detailed OWM reference with license reality, hardware, runtimes, sovereignty, fit guidance and observed change history.

20B classSee current model cardApache 2.0 + usage policy
IBM

Granite 4.2 8B

Detailed OWM reference with license reality, hardware, runtimes, sovereignty, fit guidance and observed change history.

8B-class / ~9B reported on hub131,072Apache 2.0
Moonshot AI

Kimi K2 Instruct

Detailed OWM reference with license reality, hardware, runtimes, sovereignty, fit guidance and observed change history.

1T total / 32B active128KModified MIT
Meta

Llama 4 Scout

Detailed OWM reference with license reality, hardware, runtimes, sovereignty, fit guidance and observed change history.

17B active / 109B total10MLlama 4 Community License
Mistral AI

Mistral Small 4

Detailed OWM reference with license reality, hardware, runtimes, sovereignty, fit guidance and observed change history.

119B total / 6.5B active per token256KApache 2.0
Ai2

OLMo 3 32B

Detailed OWM reference with license reality, hardware, runtimes, sovereignty, fit guidance and observed change history.

32B65,536Apache 2.0
Microsoft

Phi-4

Detailed OWM reference with license reality, hardware, runtimes, sovereignty, fit guidance and observed change history.

14B/15B class16KMIT
Qwen / Alibaba

Qwen3-30B-A3B

Detailed OWM reference with license reality, hardware, runtimes, sovereignty, fit guidance and observed change history.

30.5B total / 3.3B active32,768 native / 131,072 with YaRNApache 2.0
Qwen / Alibaba

Qwen3-32B

Detailed OWM reference with license reality, hardware, runtimes, sovereignty, fit guidance and observed change history.

33B reported on model hubUp to 131K in official Qwen3 serving examplesApache 2.0
Qwen / Alibaba

Qwen3-Coder-30B-A3B-Instruct

Detailed OWM reference with license reality, hardware, runtimes, sovereignty, fit guidance and observed change history.

30B total / ~3B active256KApache 2.0
Qwen / Alibaba

Qwen3-VL-30B-A3B-Instruct

Detailed OWM reference with license reality, hardware, runtimes, sovereignty, fit guidance and observed change history.

31B reported on model hub256K native; publisher says expandable to 1MApache 2.0
Hugging Face

SmolLM3-3B

Detailed OWM reference with license reality, hardware, runtimes, sovereignty, fit guidance and observed change history.

3B64K trained / up to 128K with YaRNApache 2.0