Quick answer
The strongest open-weight models are Kimi K3.1 Pro, GLM-5.3 and DeepSeek V4 Pro. Tencent Hy3 is the cheapest to use and has the most permissive licence (Apache-2.0). Mistral Large 4 is the leading European open model.
Open-weight models compared
Hosted price is what the lab charges through its own API; running the weights yourself costs only your hardware.
| Model | Context | Hosted price (in / out per 1M) | Licence |
|---|---|---|---|
| Kimi K3.1 Pro Moonshot, China | 1M | $2.00 / $10.00 | See the model card on Hugging Face |
| GLM-5.3 Zhipu, China | 1M | $1.40 / $4.40 | See the model card on Hugging Face |
| DeepSeek V4 Pro DeepSeek, China | 1M | $1.32 / $3.96 | See the model card on Hugging Face |
| Mistral Large 4 Mistral, Europe | 1M | $0.70 / $2.10 | Mistral Ecosystem Open Weight License |
| DeepSeek V4.1 Flash DeepSeek, China | 1M | $0.30 / $1.20 | See the model card on Hugging Face |
| MiniMax M3 MiniMax, China | 1M | $0.30 / $1.20 | MiniMax Community License (commercial use needs a licence) |
| Hy3 Tencent, China | 256K | $0.132 / $0.528 | Apache-2.0 |
Questions people ask
What is the best open-source AI model?
Kimi K3.1 Pro, GLM-5.3 and DeepSeek V4 Pro are the most capable open-weight models we track. Mistral Large 4 is the leading European option.
Can I use open-weight models commercially?
It depends on the licence. Tencent Hy3 uses Apache-2.0, which allows commercial use. MiniMax M3’s community licence requires a commercial licence, and others use their own terms, so check each model card.
Does OpenAI have open models?
Yes: gpt-oss-120b and gpt-oss-20b are open-weight under Apache 2.0, though they’re older and smaller than OpenAI’s GPT-6 models.