On 29 October 2026 Mistral announced that Mistral Large 4, a model with about one trillion parameters, is now available as open weights.[1] The company first put the model into public preview on 6 October and said then that the weights would follow at the end of the month.[2] This article reads Mistral's own pages for the release: what the model card and documentation state, which numbers are Mistral's own reporting, what the licence requires, and what it takes to run a model this size yourself. Every figure below comes from those Mistral pages, opened on 31 October 2026. We have not run the model, and no independent benchmark is cited.
What is documented
The model card and the documentation page agree on the core specification.
| Spec | Documented value | Source |
|---|---|---|
| Total parameters | About 1T (documentation: 1.05T) | Model card, documentation[3][4] |
| Active parameters per token | About 52B (49B excluding embeddings and output layers) | Model card[3] |
| Architecture | Mixture-of-experts: 128 routed experts, 4 active per token, 1 shared expert | Model card[3] |
| Modalities | Text and image input, text output; 1.6B-parameter vision encoder | Model card, documentation[3][4] |
| Context window | 1M tokens | Model card, documentation[3][4] |
| Languages | 160+ | Model card[3] |
| Reasoning effort | low, medium, high (default high) | Model card, announcement[3][1] |
| API price (per million tokens) | $0.70 input, $0.07 cached input, $2.10 output | Documentation, announcement[4][1] |
| Licence | Mistral Ecosystem Open-Weight License | Model card[3] |
Mistral also states that the model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its European datacentres.[2] The preview announcement says the model is a hybrid that combines instruction following and reasoning in one model.[2]
What is self-reported
The open-weight announcement lists a long set of results. Mistral presents them as its own: the post notes that most are Mistral's reporting, and that one preview figure is self-reported.[1] Some examples of what it claims:
- 19.7% on Cognition FrontierCode 2.0, which Mistral describes as the top open-weight score in its chart.[1]
- 41% on Terminal-Bench 4.0, up 13 points from the preview.[1]
- 51% on the Artificial Analysis Cyber Index, and a 96.6% average refusal rate on malicious cyber requests, which Mistral says is the highest it measured among open models.[1]
- A blind human-preference test in which Mistral says the model was preferred over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, with only Claude Opus 5.5 ranked higher.[1]
We have no way to check these from the pages alone. The comparison sets were chosen by Mistral, the harness settings are not described in the post, and the technical report that would document methodology is described as coming soon.[1] The practical reading is that the figures are a reason to run your own evaluation on your own tasks, not a ranking. Our guide on how to read an AI benchmark claim lists the questions to ask: who ran it, on which version, with what scaffold, and against which baselines.
One detail deserves care. The preview and open-weight posts report different numbers for the same benchmark names in places: the preview post gives 28.3% on Terminal-Bench 4.0, the later post 41%.[2][1] Mistral does not itemise the cause in the posts we read, but says at preview that the model was "still improving through ongoing reinforcement learning" at preview.[2] If you evaluated the preview, you did not evaluate the released weights.
The licence: the clause that matters
Mistral describes the Mistral Ecosystem Open-Weight License as a modified MIT licence.[1] As summarised from the licence file, it permits use, modification, distribution and further training of the weights, requires the copyright notice to be kept in copies, and adds one condition in Section 2: a licensee or affiliate that operates a Model-as-a-Service business must obtain a separate commercial licence from Mistral before commercial use.[5]
The definition is what decides who is affected. The licence defines Model-as-a-Service as giving third parties access for inference or post-training in a way that lets them have "meaningful control over the inputs, parameters, or training data."[5] Two cases are excluded from that definition: end-user products in which model capabilities are embedded in specific features, and "mere relaying of requests to models hosted by others."[5] Section 3 exempts internal use that does not make the model available to third parties, and use through Mistral's official products.[5]
What that means in practice, in our reading (this is interpretation, not legal advice):
- A company running the model for its own staff or embedding it in a product feature looks to be outside the condition.
- A cloud, API or inference provider that sells access to the weights is the case the condition is aimed at.
- A grey area exists between "embedded feature" and "meaningful control over inputs and parameters". If your product exposes prompting, fine-tuning or parameter control to customers, read Section 2 and its definition closely, and ask Mistral or counsel.
For a wider comparison of how this kind of condition differs from Apache 2.0 or the Llama licence, see open-weight model licences compared and the explainer on open weights versus open source. The extra Model-as-a-Service condition is why we call this model open weight rather than open source.
What it takes to run it yourself
Mistral's model card says the weights are on Hugging Face, with serving recipes for vLLM (recommended) and SGLang, plus a Transformers path installed from main.[3] It also lists a separate NVFP4 checkpoint for Blackwell GPUs and GGUF builds for llama.cpp.[3] The card gives no explicit memory requirement, and its examples use an 8-GPU node with tensor-parallel size 8.[3]
A rough bound is simple arithmetic, ours rather than Mistral's. A mixture-of-experts model computes with only its active parameters per token (about 52B here), but the router can send any token to any expert, so all roughly 1.05T parameters must be resident in memory. At 2 bytes per parameter that is about 2.1 TB for weights alone; at 4 bits per parameter it is about 0.5 TB. Both ignore the key-value cache, which grows with context length and concurrent requests, and the real checkpoint sizes are not stated on the pages we read. The same reasoning is behind the memory-bound analysis in GPU vs TPU for AI: active parameters set how much arithmetic each token needs, but total parameters set how much hardware you must rent.
The consequence: for most teams the economic comparison is not "free weights versus paid API". It is the API price in the table above against the cost of keeping a multi-GPU node busy enough to beat it. Cached input at $0.07 per million tokens is ten times cheaper than uncached input, which favours workloads with long repeated prefixes.[4]
A decision guide
| If you are... | Consider | Check first |
|---|---|---|
| A team that wants the model with no infrastructure | Mistral's API at the listed prices | Data residency terms in your own contract |
| An enterprise with data that cannot leave your network | Self-hosting the weights | Hardware budget, serving stack, the licence's internal-use exemption |
| A cloud or inference provider | A commercial licence from Mistral | Section 2 of the licence and the MaaS definition |
| Comparing models for coding or agents | Your own evaluation on your tasks | Whether you tested the preview or the released weights |
What remains uncertain
- The technical report, with training and post-training methodology, is not yet published.[1]
- Mistral says it sees "a potential for uplift to conduct real attacks" and is changing its guardrails; a restricted cyber variant is in preview for vetted partners only.[1] Our interpretation: refusal-rate figures describe the released checkpoint, and a fine-tuned derivative of open weights may behave differently.
- No independent evaluation of the released weights was found in this run.
- The announcement page carries no date; we use 29 October 2026 from the date Mistral's news listing shows for the post.
Sources and method
We opened Mistral's open-weight announcement, its preview announcement, the Hugging Face model card, the model documentation page and the licence file on 31 October 2026, and quote only what those pages state. Licence provisions are summarised, not quoted in full, and the licence text should be read directly before any commercial decision. The memory arithmetic is ours and labelled as such.
