Mistral has published the weights of Large 4, a roughly 1 trillion-parameter model, under a modified MIT license: you can use and modify it freely, but a company that sells it as a paid hosted service needs a separate license. For most teams the realistic way in is still Mistral’s API, because self-hosting needs a multi-GPU server. This is an analysis of the announcement and coverage, not a hands-on test.

What was announced
Mistral’s own Large 4 post, dated October 6, 2026, introduced the model in public preview through the API and promised the weights later: it describes a 1 trillion-parameter natively multimodal model with 52 billion active parameters, and says the weights drop at the end of the month. According to Online Tech Tips, Mistral released the full downloadable weights on October 29, 2026. DataNorth (October 30) also reports that the weights are on Hugging Face.
The license in plain terms
Both outlets say Large 4 uses the Mistral Ecosystem Open-Weight License, which Mistral describes as a modified MIT license. Internal use carries no license fee. The catch, as reported, is that businesses wanting to offer the model itself as a paid hosted service need a separate commercial license. Coverage says the previous generation, Large 3, used Apache 2.0, so this is a tighter grant than before. I have not read the license text itself; check it before building a product on the model.
Key facts

| Item | Detail |
|---|---|
| API preview | October 6, 2026 (Mistral) |
| Weights released | October 29, 2026 (secondary coverage) |
| Size | About 1 trillion parameters, 52 billion active (Hugging Face lists about 49 billion) |
| API price (USD, per million tokens) | $0.70 input, $2.10 output, $0.07 cached input |
| License | Mistral Ecosystem Open-Weight License |
Who it matters for
- Companies with data-residency or privacy rules: downloadable weights make self-hosting possible, which an API-only model cannot offer.
- Startups reselling model access: the hosting restriction is the clause to read first.
- Individual developers: this is not laptop territory. DataNorth says even a compressed version needs a server with four Nvidia B200 or B300 GPUs, and Online Tech Tips cites a deployment recipe that runs the FP8 version across eight NVIDIA H200 GPUs. Most people will use the API.
Limits and open questions
The benchmark numbers I have seen are self-reported by Mistral, and I found no independent evaluation yet. There is also a safety trade-off: DataNorth notes that weights you download today will not receive Mistral’s later safeguard changes unless you download a new version, so self-hosters must set their own controls over users, logging and tool access. Reports differ slightly on active parameter counts and the exact release day, which is why this post cites each one to its source.
My take
Open weights at this scale are useful mainly to organizations that already run serious GPU infrastructure. For everyone else, the practical comparison is API price and quality, the same lens we used for OpenAI’s GPT-6.1 Terra pricing. If you are weighing where to run AI automation, our look at Claude on Google Cloud covers the managed-cloud side, and the comparison checklist helps structure a trial.
Sources
- Mistral: Large 4, October 6, 2026.
- Online Tech Tips: Mistral releases Large 4 open-weight model, early November 2026.
- DataNorth: Mistral Large 4 open weights, October 30, 2026.

