Mistral Large 4 is live in preview. Weights wait until month end.

Mistral put Large 4 into a public preview API on 6 October 2026. The internal name is Le Chonk. The model id on the docs page is mistral-large-4. Weights are not in the download folder yet. Mistral says those land at the end of the month. VentureBeat reported 27 October as the date given to reporters. Until then this is an API product, not a file you can pin to your own cluster.

I read the launch as a founder who already pays for one European API and one US API. The useful question is not whether a trillion-parameter badge is impressive. It is what you can call today, what it costs for the next two weeks, and what you still cannot self-host.

What actually shipped

The blog post calls it a 1 trillion-parameter natively multimodal model with 49 billion active parameters. The model card, dated the same day, says 1.05 trillion total, 52 billion active, plus a 1.6 billion parameter vision encoder. Granular mixture-of-experts. Text and image in, text out. No audio, no video, no image generation.

Context on the model card is 1 million tokens. Artificial Analysis, writing the same day, listed 512k. Size a production prompt against a live call, not against either sentence, until Mistral reconciles the two pages.

The preview is served from the same European datacenters that trained it. Mistral says training ran from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own facilities. A large share of the training mix covered more than 160 languages, including every official language of the European Union. The company also says a European deployment is operated end to end by Mistral, under European law, separate from other cloud providers. That is the sovereignty pitch, and why a regulated buyer might trial this before waiting for weights.

The docs page lists chat completions, function calling, structured outputs, document Q&A, prefix completions, batch, and the agents and conversations endpoints, plus built-in tools. Studio is where they send people who want to click before they wire a key.

The price that expires

List price on the model card is $1.36 per million input tokens, $0.14 per million cached input tokens, and $4.18 per million output tokens. The changelog for 6 October says launch pricing is 50% off for two weeks. The struck-through numbers on the same card match that: $0.68 input, $0.07 cached input, $2.09 output.

Two weeks from 6 October is 20 October. After that, assume the higher line unless the card still shows the discount. A million output tokens at the promo rate is $2.09. At list it is $4.18. A long agent trace that writes 200,000 output tokens is about $0.42 during the promo and about $0.84 after. A 100,000-token document at the promo input rate is under 7 cents before cache.

That is cheap next to the closed frontier names Mistral keeps citing. It is not free, and preview models move. Budget a cap. Batch is listed, so overnight evals do not have to sit on the interactive tier.

Scores they published, and one outside check

These numbers are from Mistral's 6 October post unless I say otherwise. Treat them as vendor evals.

Cyber is the section they lead with. On one Artificial Analysis Cyber Index test that asks a model to reproduce a real vulnerability in open-source software and then patch it, Mistral reports 82%, and calls that the highest of any model on that test. Cybench, 40 exercises from security competitions: 93% solved. They also say several closed models, including Claude Opus 5.5 and GPT-6 Astra, score near zero on the same reproduce-and-patch test because they refuse the task. Artificial Analysis's own CyberGym-E2E-AA board, a filtered set of 131 memory-safety tasks, lists Mistral Large 4 Preview at 81.7% pass@1, ahead of MiMo-V2.6-Pro at 78.6% and GPT-6 Luna (Max) at 77.9%. That outside number sits close to the company figure. A defender who needs a model to confirm a flaw exists will not get that from a system that declines the prompt. Mistral is also red-teaming with cybersecurity partners on a build with reduced moderation. The public preview is not that build.

Coding, still from the post: 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, 28.3% on Terminal-Bench 4. Combined Coding Agent Index 49.8%, which they place ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max. A blind Surge AI rating of coding quality, 1 to 5, identities hidden: ML4 Preview at 3.74, behind Claude Opus 5 at 4.22, ahead of GLM-5.3 at 3.60, Kimi K3 at 3.59, and GLM-5.2 at 3.40. Second of five is a real result. It is not a claim that this replaces the model your senior engineers already trust on hard diffs.

Agents: 59.9% on AutomationBench, 657 business workflows across Gmail, Google Sheets, Slack, and Salesforce, ahead of Kimi K3, MiMo-V2.6-Pro, and DeepSeek V4 Pro on their chart. AA-Briefcase, long-horizon knowledge work, 1,393 Elo, ahead of DeepSeek V4 Pro. Visual grounding on Dense 200: 42% versus 41% for GPT-6 Astra. That one-point gap is the line about beating a closed frontier model. It is a narrow task. Do not stretch it.

Legal and finance: Mistral says third-party evals through vals.ai had Large 4 ahead of GPT-6 Astra on the tasks they ran, and ahead of other open models on Harvey's Legal Agent benchmark. I have not re-run those. Safety: 93.3% attack resistance on Lakera's public B3 set, and 1.691 out of 2 on KORABench. They also say the average refusal rate on cyber prompts drawn from JailbreakBench, StrongREJECT, and AgentHarm is higher than the other open models they compared. Strong cyber scores and a higher refusal rate can both be true. Read your own prompts before you put this on a security workflow.

What you cannot do yet

You cannot self-host it. The technical documentation for downstream providers, version 1, published 6 October, says it will be updated when weights ship. License is not on that PDF yet. VentureBeat said the expectation is a custom Mistral license, not a standard permissive grant. Dropping weights on a machine you already pay for starts at the end of October at the earliest, and the license text is part of the decision.

The same PDF lists a maximum of 8 images. Artificial Analysis wrote that the API now accepts 100 images per request, up from 8 on earlier Mistral models. Another page conflict. If the product is a drawing set or a satellite tile stack, send a probe before you promise a gallery limit.

Preview also means the weights can still move. A score you log this week may not match the model id next week even if the string stays mistral-large-4. Pin evals to a date. There is no image output. If you needed a generator, this is the wrong card.

How I would trial it this week

I would not migrate a production agent on a preview. I would spend the promo window on three checks.

First, a cost replay. Take 50 real traces from the model you use now. Send the same tool schema to mistral-large-4. Log input tokens, cached input, and output tokens. At $0.68 / $0.07 / $2.09 you can see whether the bill drops before you care about the leaderboard.

Second, a refusal check on the work you actually do. Contract markup, spreadsheet repair, or a coding ticket with a failing test is the easy pass. Vulnerability confirmation belongs on a repo you own, with refusals logged apart from wrong answers. A near-zero score from a closed model on a cyber bench is often a policy result, not a skill result.

Third, a long-context check. Stuff a real document pack to 200k, then to whatever the API accepts above that. If the card's 1 million holds, say so in your notes. If calls die near 512k, believe the call.

Region matters if a customer contract names a jurisdiction. Ask which region the preview key is pinned to before you paste client data. The European-operated deployment is the line in the post. It is not automatically the region behind every new key.

Weights at month end change the math again. Self-host only pays if the license allows the use and the footprint fits hardware you can rent. Until that PDF updates, the product is the preview API, half price until about 20 October, model id mistral-large-4.