Mistral Large 4: What It Is, What It Costs, and the One Line Still Blank on Mistral's Own Card

Mistral Large 4 is Mistral AI's flagship generalist model and the top of the Large line that Mistral Large 3 held before it — a natively multimodal mixture-of-experts model that Mistral's own documentation describes as having 1.05 trillion total parameters with 52 billion active, a 1.6-billion-parameter vision encoder, a 1M-token context window, and a rate-card price of $1.36 per million input tokens against $4.18 per million output. The API name is mistral-large-4. What is not on that card is the thing a large share of readers arrive looking for: the licence badge on Mistral's own model page reads "Open" and then names nothing at all, and there is no checkpoint repository to download. This is the description of the model as it stands on 13 October 2026.

The framing matters, so it goes up front. Mistral announced the public preview of Mistral Large 4 on 6 October 2026, and this blog already carries that news piece plus the nine head-to-head pages that follow from it. What has been missing is the plain page underneath them. That gap is measured rather than assumed: the bare name drew 3,141 impressions and 68 clicks on our own property in the twenty-eight days to 13 October, spread across five queries, while the qualified rows below it — artificial analysis at 129 impressions, preview at 122, benchmark at 84, ollama at 70 — are all small. The comparative rows, meanwhile, convert several times better than the bare term does, which is precisely why the comparison pages were written first. This is the page for the reader who typed the name and wanted to know what the thing is.

The name, and the 4-0 sitting in our own URLs

Mistral writes the model as Mistral Large 4. Its launch post offers two informal handles alongside that — "ML4", and "le Chonk" — and its documentation serves the model card from a page reachable under both mistral-large-4 and mistral-large-4-0. The badge on the card itself reads mistral-large-4, and the model selection guide lists mistral-large-4 as the API name, so the plain form is the vendor's current one and the trailing zero is a legacy alias carried from the launch-day link.

That is worth one sentence of reconciliation because every page already published here uses the 4-0 form — the preview write-up, the hosting page and all nine comparisons are filed under mistral-large-4-0-*. Those URLs resolve to the same model this page describes. This page simply takes the vendor's spelling for the reference entry, so a reader who lands here from the bare name meets the model under the name its maker currently prints.

The three status labels on Mistral's card are worth reading together, because they are the model's whole situation in three words: Public Preview, Open, and v26.10. The first says the serving configuration can still move. The second is the open-weight claim. The third is the revision, and it is the only one of the three that is simply a fact.

The specification, as Mistral states it

Everything in the list below is Mistral's own wording from its model card, its model selection guide and its launch post. Where an independent lab reads the same number differently, that reading is named on the line rather than averaged in.

• Parameters — 1.05 trillion total with 52 billion active parameters, on a granular sparse mixture-of-experts architecture. The launch post rounds the total to "1 trillion-parameter"; the model card gives the 1.05T figure. Artificial Analysis records the active count as 49 billion rather than 52 billion, which is a three-billion gap on the same dimension.

• Vision encoder — 1.6 billion parameters, on a model Mistral describes as natively multimodal rather than adapted to images after the fact.

• Modalities — text and image in, text out, with reasoning included. Mistral's selection guide classes the output as "reasoning, text", and Artificial Analysis records it as a reasoning model with a note that a non-reasoning variant may also exist. The launch post's own product card calls it a "hybrid instruct-and-reasoning MoE".

• Context window — 1M tokens on Mistral's card. Artificial Analysis reads 524,288 tokens on the configuration it tests, which is the difference between a vendor ceiling and a served window.

• Price — $1.36 per million input tokens, $0.14 cached input, $4.18 per million output, per Mistral's published pricing page. That page also carries a launch promotion showing the same three lines struck through at $0.68, $0.07 and $2.09.

• Languages — more than 160 languages in the training data, including every official language of the European Union, on Mistral's own account.

• Licence — unnamed. More on that below, because it is the single most load-bearing blank on the page.

Licensing and weights: what this release actually publishes

Mistral Large 3 shipped under Apache 2.0. Its model card on Hugging Face carries the licence tag in the metadata, and its weights are downloadable in FP8, NVFP4 and BF16. Mistral Small 4 and the Ministral 3 sizes are Apache 2.0 as well, and Mistral Medium 3.5 is listed as Modified MIT. None of that transfers. A licence is read off the release in front of you, and on Mistral Large 4 the vendor has not yet written one down.

Three places on Mistral's own surfaces show the blank, and a reader can check each of them in a browser:

• The model card at docs.mistral.ai carries a licence chip that renders the single word "Open" with no licence name attached to it.

• The model selection guide — the side-by-side table Mistral publishes for developers choosing between its models — has a Licence column, and every neighbouring row is populated: Apache 2.0 for Mistral Small 4, Modified MIT for Mistral Medium 3.5, Apache 2.0 for Mistral Large 3. The Mistral Large 4 row's Licence cell is empty.

• The weights tab on the model card reads "Coming soon".

Artificial Analysis, which tracks the same field independently, records the licence name and licence URL for Mistral Large 4 as null, marks the model isOpenWeights: false, and categorises it as proprietary. That flag is not a contradiction of Mistral's marketing; it is a statement about what exists today. A preview served from the vendor's own cluster, with no repository and no licence text, is a callable model rather than an open-weights release.

What Mistral has committed to is a date: "We will release the weights by the end of the month," in the launch post's own words, alongside further detail on architecture, additional benchmarks and post-training methodology. Ahead of that, the company says it is running a red-teaming window with vetted partners and national authorities. The organisation's Hugging Face account, read on 13 October, still has nothing newer than 8 October, and neither entry is a Large-line checkpoint.

So the accurate description of the open-weight claim today is that it is a plan with a deadline attached to it, made by a lab that has shipped Apache 2.0 weights at this tier before. That is a stronger position than a vague promise, and it is a weaker one than a download link. The distinction between those two things is the entire reason this section exists.

Where Mistral Large 4 runs, in Mistral's words

Mistral says the preview API is available on Mistral Studio, and that the model is "available across multiple regions worldwide, including a European deployment that Mistral operates end-to-end, independently of other digital service providers and under European law". The company also says the preview is served on the same infrastructure the model was trained on — 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters — and that the model will be able to run on private cloud or on-premise for the workloads it is aimed at. Those are the vendor's statements about where its model runs, attributed to the vendor and not to us.

Separately and plainly: Mistral Large 4 is not in OrcaRouter's catalogue. Our live model list, read on 13 October, serves 207 model identifiers and not one of them is a Mistral-family model, under mistral-large-4, mistral-large-4-0 or any other spelling. No route is being claimed here, and nothing in this paragraph should be read as one.

One deployment question that this page will not re-answer is the Ollama listing. Mistral Large 4 appears in that library as a single cloud tag rather than a downloadable size ladder, which is a distinct story with its own write-up — mistral-large-4-ollama-cloud-only — and a reader who arrived wanting to know whether they can pull the weights onto a local machine should read that page rather than this one.

The benchmark numbers, and who ran them

Mistral's launch post reports a dense set of scores, and it is unusually explicit about which of them the company did not run. On the agentic-coding side it lists 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4.0, combining into a 49.8% Coding Agent Index score that the post says places the model ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max. On agentic workflows it reports 59.9% on AutomationBench across 657 business workflows, and on long-horizon knowledge work it reports 1,393 Elo on AA-Briefcase. On its cyber evaluations it reports a top-five position on the Artificial Analysis Cyber Index, 82% on a reproduce-and-patch test described as the highest of any model, and 93% on Cybench's 40 competition exercises. On multimodal grounding it reports 42% on Dense 200 against 41% for GPT-6 Astra. All of those are vendor-reported figures; some are described in the post as evaluated by a third party ahead of that harness going public, which is a stronger provenance than a self-run number and still not a published one.

The independent numbers are thinner and more useful. Artificial Analysis scores Mistral Large 4 Preview at 38.4 on its Intelligence Index v4.3, which places the model above the median of its comparison class and 65th of 216 models on that index. The same page puts the cost of running one Intelligence Index task at $1.13, and records something the launch post does not mention at all: the model generated 200 million output tokens during the index run against a median of 85 million, ranking 107th of 114 on verbosity. A model that is above average on quality and far above average on tokens spent per answer changes the arithmetic of a bill in a way the headline scores do not.

There is also one blind evaluation in the launch post that is worth separating out from the rest, because the method is different. Surge AI ran a blind human comparison on coding quality with model identities hidden; professional annotators rated Mistral Large 4 Preview 3.74 on a 1-to-5 scale, second of five, ahead of Kimi K3 at 3.59 and GLM-5.3 at 3.60 and behind Claude Opus 5 at 4.22. A blind panel with a named third party at the other end is a different class of evidence from a score the vendor selected, and it is the one figure in the announcement that a sceptical reader should weight most heavily.

What changed from Mistral Large 3, in three sentences

Mistral Large 3 is a 675-billion-parameter sparse mixture-of-experts model with 41 billion active parameters, a 256k context window, a 2.5-billion-parameter vision encoder, and Apache 2.0 weights you can download today. Mistral Large 4 raises the totals to 1.05 trillion with 52 billion active on the vendor's card, lifts the stated context to 1M tokens, and adds reasoning to a model Mistral previously described as an instruct-only multimodal — which is a change of kind, not only of scale. It also arrives, for now, without the licence and the checkpoint that made Large 3 straightforward to adopt.

Artificial Analysis records Mistral Large 3 as deprecated and points its deprecation target at Mistral Large 4, which is the clearest signal available that these are successive generations of the same slot rather than parallel options. Mistral's own pricing page agrees: Large 3 sits at $0.50 input and $1.50 output per million tokens, and Large 4 is listed above it with the promotional pair in place.

The comparison set, as an index rather than a verdict

Nine head-to-head pages already exist for this model, and this page deliberately does not re-run any of them. For a reader deciding between Mistral Large 4 and something else, the pages are mistral-large-4-0-vs-claude-opus-5, -vs-deepseek-v4-pro, -vs-gemini-3-1-pro, -vs-glm-5-5, -vs-gpt-6-sol, -vs-grok-4-7, -vs-kimi-k3, -vs-minimax-m3 and -vs-qwen-3-8-max. The measured demand sits in exactly those pages — mistral large 4 vs claude at 78 impressions and 11 clicks, vs opus 5.5 at 84 and 10 — and both convert well above the bare term's rate, which is an argument for keeping the definition here and the adjudication there.

The dependency worth naming is mistral-large-4-0-vs-glm-5-5, because GLM-5.5 is itself a page in this same batch and the two write-ups have to agree about a comparison rather than argue past each other. Everything else on the list is a stable target.

What is not settled

Whether the weights land by the end of October, and under what licence. And whether the promotional pricing survives the preview window, since the standard rate card and the launch rate differ by exactly half on all three lines. Both are checkable by a reader without asking anyone: the licence and the checkpoint will appear on Hugging Face or they will not, and the rate card will either still show a struck-through number or it will not.

Until then the posture this model invites is the one its own stage label suggests. Call it, measure it against your own workload — including the token spend, which is the number the launch post leaves out — and keep the production path on something whose behaviour is not scheduled to change. A preview is not merely a label on a price sheet. It means the model, the licence, and the bill can all move at once.

At OrcaRouter we route more than 200 models through one OpenAI-compatible endpoint at the provider's list price with 0% markup passed through — so a vendor price change is a price change on our side the same day, whichever model it lands on. Automatic failover is the other half of that for anything at preview stage: it lets a slice of real traffic meet an unproven model while a settled one holds the rest, so a regression routes around instead of paging someone. Neither of those is a claim about Mistral Large 4, which we do not serve. They are the shape of the answer for the next model, and this one, that ships before its licence does.

None of that is hypothetical: the documented AI failures and agent incidents this reference is written against are collected in the OrcaRouter AI Incident Archive , where each entry carries its own sources, severity, confidence grade and disputed flag.