Mistral’s open weights widen deployment choices, with limits still to test
Mistral Large 4 can now be downloaded for self-hosting. Independent benchmark results add evidence of capability, while deployment terms and real-world reliability remain separate questions.
Mistral’s 29 October release of the improved Large 4 model gives organizations a new choice: run downloaded weights on their own infrastructure instead of relying on a hosted model service. The company had offered a public preview earlier in October. The change concerns control over deployment as well as the capabilities Mistral says it has improved. [1][2]
Independent testing supplies a narrower basis for judging those capabilities. Artificial Analysis lists a score of 26 for the High reasoning version on its Intelligence Index, compared with a median of 23 among comparable open-weight models. That supports an above-median result on that evaluation; it does not establish that the model is best for every task. EWN has not run its own model tests. [2]
Availability also has conditions. Mistral’s published license permits modification and deployment, but includes a separate commercial-license requirement for certain Model-as-a-Service businesses, alongside stated exceptions. Downloadable weights therefore should not be mistaken for an absence of use conditions. This is a description of the published terms, not a determination of how they apply to a particular business. [3]
Taken together, the release and independent results separate three questions that marketing can blur: where a model can run, what a benchmark measures, and whether it performs reliably in a particular workflow. The release changes the first; the independent index informs the second. Neither source settles the third. Mistral says a full technical report is forthcoming, without giving a publication date in the announcement reviewed here. [1][2]
Sources & context · 3
- Mistral
Original release announcement: downloadable improved weights after the October preview and a forthcoming technical report. Vendor capability claims are not treated as independent validation.
- Artificial Analysis
Independent measurement page for the High reasoning version: Intelligence Index 26 and comparable-model median 23. This is a benchmark result, not an EWN test or a production-reliability guarantee. The page identifies 29 October as release date.
- Mistral on Hugging Face
Original model license, including commercial-service conditions and exceptions; the same originating organization as source 1, not a third independent contribution.