Mistral Large 4 Weights Shipped Under a Modified MIT. Host It Yourself for Free; Host It as a Service and You Need Mistral's Separate License.
Mistral released the open weights for Large 4 on October 29 on Hugging Face, three weeks after the API-only preview. The docs say 1.05 trillion total parameters, 52 billion active, and a 1 million token context. The license is the Mistral Ecosystem Open-Weight License, a modified MIT that allows use, modification, fine-tuning, and commercial products, except that running the model as a service for third parties requires a separate commercial license. API list price is $0.70 in, $0.07 cached, $2.10 out per million tokens. Artificial Analysis scores it 26 at high and 23 at medium on Intelligence Index v5. Mistral's coding, cyber, and refusal numbers are its own. We read the release post, the docs, the license, and AA's model page.
By T.W. Ghost
Mistral released the weights for Mistral Large 4 today, October 29, 2026, at mistralai/Mistral-Large-4-1T-A52B on Hugging Face. When we covered the preview on October 6, open weights were a promise for the end of the month with no license named; two days ago the placeholder still showed a countdown. The files are now there, and so is the license, and the license is the part to read.
It is called the Mistral Ecosystem Open-Weight License, and it is a modified MIT. You may use, modify, fine-tune, and build commercial products on the model. What you may not do without a separate commercial license from Mistral is operate it as a Model as a Service, which is to say sell hosted inference or hosted post-training of this model to third parties. Self-host for your own product, fine; host it for other people's products, call Mistral. That clause makes "open weights" true and "open" in the usual hosting-provider sense false, which is the kind of distinction this site exists to print. We read the release post, the model docs, the LICENSE file on Hugging Face, and Artificial Analysis's model page. Here is what held up.
What You Are Downloading
| Mistral Large 4 (open weights, October 29, 2026) | |
|---|---|
| Repository | mistralai/Mistral-Large-4-1T-A52B on Hugging Face |
| Size | 1.05T total parameters; 52B active; 1.6B vision encoder separate (Mistral docs) |
| Context | 1M tokens |
| License | Mistral Ecosystem Open-Weight License, modified MIT; Model-as-a-Service operators need a separate commercial license |
| Reasoning modes | Low, Medium, High (vendor) |
| API list price | $0.70 in / $0.07 cached / $2.10 out per 1M tokens (release post and docs) |
| AA Intelligence Index v5 | 26 (high); 23 (medium) |
| Pentest variant | Partner-only (vendor) |
A52B means 52 billion active, in the docs and in the repo name. Our launch post tracked a 49-billion figure that moved to 52 on Mistral's blog; a 49B-excluding-embeddings reading still circulates. The docs say 52B active, and we print the docs.
The API price in the docs is not the preview price. Our October 6 post listed the preview at $1.36 in and $4.18 out, described as half price for two weeks. The release post and model docs now list $0.70, $0.07 cached, and $2.10. Mistral's consumer pricing page is for its Vibe plans and does not carry this table; take the price from the docs.
The index score is on v5. Artificial Analysis rebuilt its index on October 27, and 26 at high is a v5 number. The 38 in our launch post was v4.3.2. They are not comparable, and neither is the "most intelligent model outside the US and China" line AA wrote on the old scale.
The Benchmarks, Sorted by Who Reported Them
| Claim | Number | Who |
|---|---|---|
| Intelligence Index v5 | 26 (high), 23 (medium) | Independent (Artificial Analysis) |
| FrontierCode 2.0 | 19.7% | Vendor (Mistral; private Cognition benchmark) |
| Terminal-Bench 4.0 | 41% | Vendor |
| Coding Agent Index | 48, tied with GLM-5.3 at 47% fewer tokens | Vendor (Mistral's citation) |
| Cyber | 51% | Vendor |
| Refusal rate | 96.6% | Vendor |
| "Best open coder"; "eighth among open weights"; "best non-Chinese open model" | rank claims | Vendor and Press (a Trending Topics table, not an AA primary) |
Three honest observations:
- The coding story is Mistral's and it runs through a private benchmark. FrontierCode is Cognition's held-out test; Mistral quoting a score on it is a vendor citing a number only Cognition can check.
- The one independent number is 26. That is where AA puts Large 4 on its new scale. The coding and cyber percentages above come from Mistral's release post.
- "Eighth open" is a press table. Until AA's filtered open-weight board is quoted directly, the rank is secondary.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| Weights released Oct 29, 2026; Hugging Face repo | Verified (Mistral release post; repository; Mistral's legal lifecycle page) |
| 1.05T total; 52B active; 1.6B vision encoder; 1M context | Verified (Mistral docs) |
| Modified MIT with a Model-as-a-Service commercial-license clause | Verified (LICENSE file; release post) |
| API $0.70 / $0.07 cached / $2.10 | Verified (release post; docs) |
| AA v5 26 high, 23 medium | Verified (AA model page) |
| FrontierCode 19.7%; Terminal-Bench 4.0 41%; Coding Agent 48 tied GLM-5.3; Cyber 51%; refusal 96.6% | Vendor |
| "49B active" | Unverified; contradicted by the docs' 52B |
| "Eighth open" / best non-Chinese | Press |
| AI Act downstream-provider documentation updated after the ship | Unverified (was proprietary language before the release) |
Who Should Use It
Download it if you want a trillion-parameter open model you can self-host for your own product under a license that lets you do that, and you have the hardware for 1.05T total parameters.
Skip it, or call Mistral first, if your business is hosting models for other people; the clause is in section 2 of the license and it is the whole reason the word "modified" is in front of MIT. Skip the coding ranking until someone other than Mistral runs it.
The frontier is elsewhere: on v5 the released open weights that lead are at 32 and Large 4 is at 26, and the US open models we covered this month, Muse Spark Open and Reflection Beam, have not been re-scored on the new scale in anything we read. Our model comparison has the current standings side by side.
Sources
- Mistral: Mistral Large 4 open-source release
- Mistral docs: Mistral Large 4
- Hugging Face: mistralai/Mistral-Large-4-1T-A52B LICENSE
- Artificial Analysis: Mistral Large 4 model page
- Trending Topics: Mistral Large 4 open weights and the Artificial Analysis ranking (press)
*Not sure whether a self-host-only open license fits how you plan to serve a model? Take the free 2-minute quiz and get matched. Then read the launch post this one updates, with the preview price and the old scale: Mistral Large 4.*