Mistral Large 4 weights are out: open-weight coding at half the tokens, with a no-hosting clause
Mistral kept its end-of-October promise. On October 29 it put the weights of Mistral Large 4, nicknamed Le Chonk, on Hugging Face: about 1 trillion parameters, 52 billion active per token, text and image in. The Hacker News thread passed 1,700 points the same day.
Mistral's pitch is efficiency, not raw score. On the Artificial Analysis Coding Agent Index it scores 48%, which Mistral calls "effectively tied" with GLM-5.3, the best open model, "while using 47% fewer tokens per task." On Terminal-Bench 4.0 it reached 41%, up 13 points from the preview three weeks ago, with 30% fewer output tokens. On Cognition's private FrontierCode 2.0, where the grade is whether a maintainer would merge the change, it leads open models at 19.7%. The API price is $0.70 per million input tokens, $0.07 cached and $2.10 output. When agent bills are dominated by tokens per task, halving tokens at the same score is worth more than two extra benchmark points.
The fine print is the license. Large 3 shipped under Apache 2.0. Large 4 ships under a new "Mistral Ecosystem Open-Weight License," which Mistral describes as a modified MIT: free to use, modify, fine-tune and commercialize, with one exception. Any business running a model-as-a-service offering needs a separate commercial license from Mistral. Self-host it inside your company, fine. Sell it as an API, call Mistral first. GLM-5.3 and MiniMax M3.1 Flash went the same way in recent weeks. "Open weights" and "open source" are separating fast, and the line is drawn exactly at the inference providers.
The cyber section is also worth a look. Mistral says the model handles defensive tasks without needless refusals and refuses 96.6% of malicious cyber prompts. It also admits that live traffic since the preview showed "a potential for uplift to conduct real attacks," and that guardrails are being tightened after release. A specialized pentesting variant goes only to vetted partners. Once weights are public, though, those guardrails are only as strong as whoever downloads them.
Our read: this is the strongest Western open-weight release of the year, and it competes on cost per task. Teams building agents on self-hosted models get a real option. Hosting companies get a phone call to make.
Announcement: https://mistral.ai/news/mistral-large-4-oss/ Weights: https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B
← Back to all articles
Mistral's pitch is efficiency, not raw score. On the Artificial Analysis Coding Agent Index it scores 48%, which Mistral calls "effectively tied" with GLM-5.3, the best open model, "while using 47% fewer tokens per task." On Terminal-Bench 4.0 it reached 41%, up 13 points from the preview three weeks ago, with 30% fewer output tokens. On Cognition's private FrontierCode 2.0, where the grade is whether a maintainer would merge the change, it leads open models at 19.7%. The API price is $0.70 per million input tokens, $0.07 cached and $2.10 output. When agent bills are dominated by tokens per task, halving tokens at the same score is worth more than two extra benchmark points.
The fine print is the license. Large 3 shipped under Apache 2.0. Large 4 ships under a new "Mistral Ecosystem Open-Weight License," which Mistral describes as a modified MIT: free to use, modify, fine-tune and commercialize, with one exception. Any business running a model-as-a-service offering needs a separate commercial license from Mistral. Self-host it inside your company, fine. Sell it as an API, call Mistral first. GLM-5.3 and MiniMax M3.1 Flash went the same way in recent weeks. "Open weights" and "open source" are separating fast, and the line is drawn exactly at the inference providers.
The cyber section is also worth a look. Mistral says the model handles defensive tasks without needless refusals and refuses 96.6% of malicious cyber prompts. It also admits that live traffic since the preview showed "a potential for uplift to conduct real attacks," and that guardrails are being tightened after release. A specialized pentesting variant goes only to vetted partners. Once weights are public, though, those guardrails are only as strong as whoever downloads them.
Our read: this is the strongest Western open-weight release of the year, and it competes on cost per task. Teams building agents on self-hosted models get a real option. Hosting companies get a phone call to make.
Announcement: https://mistral.ai/news/mistral-large-4-oss/ Weights: https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B
Comments