Overview
Mistral Large 4 is the frontier model Mistral released on October 6, 2026. It is a sparse model with 1T total parameters and 49B active, natively multimodal, with a 512K context window and up to 256K output tokens. It currently ships as a Research Public Preview on Mistral's own API and on OpenRouter, with the 1T weights planned for release at the end of October. That means it is not strictly an open model yet, but a staged rollout where you can build against it now and get the weights a few weeks later.
The weight of the release shows up in the scores. Artificial Analysis puts it at 38 on the Intelligence Index, level with GPT-6 Luna at max effort and just behind DeepSeek V4.1 Flash at 39, which makes it the highest scoring model from outside the United States and China and puts France back in that position. On the coding side it entered the Code Arena WebDev leaderboard at 1534, ranked 45th, some 304 points above Mistral Large 3.
Key Features
- 1T parameters, 49B active: A sparse architecture with a trillion total parameters and 49 billion active per token, trading a very large model capacity for manageable per-token compute. This is what lets it sit next to the frontier tier on intelligence while staying serviceable on cost
- Natively multimodal with 100 images per request: Text and image in, text out. The API now accepts 100 images per request where previous Mistral models took 8, and the effect shows directly in document understanding: 19 percent on GDP.pdf, an 18-point gain over Mistral Large 3
- 512K context, 256K output: A 512K token context window with up to 256K output tokens in one go, enough for long repositories, long documents and long sessions
- Cyber capability in the top tier: It scores 50 on the Artificial Analysis Cyber Index, level with GLM-5.3-Flash, and once weights land it will rank among the top three open weights models on that index. Its strongest result is 82 percent on CyberGym-E2E-AA, ahead of MiMo-V2.6-Pro at 79 percent and GPT-6 Luna at max effort with 78 percent
- Verifiable on coding and agent arenas: It is live in Arena's Agent Arena and in the WebDev, Text and Vision tracks of Code Arena. On WebDev it sits at 1534 in 45th place, a large jump from Mistral Large 3 in 130th and above Mistral Medium 3.5
- Weights coming at the end of October: The preview runs on Mistral's API and OpenRouter. Mistral plans to publish the 1T parameter weights at the end of October, after which it moves from a hosted service to an open weights model you can deploy yourself
Use Cases
- European teams with data residency or supplier jurisdiction requirements who need a frontier model they can run locally
- Enterprise document processing that needs long context plus many images, such as submitting hundreds of scanned pages and charts at once
- Security automation, red teaming and defensive evaluation where the Cyber Index and CyberGym results matter
- Teams planning self-hosting once weights land, who can use the preview window to wire up the pipeline and tune prompts
Pros
- 38 on the Intelligence Index, the highest score from outside the US and China
- 512K context with 256K output handles long jobs in a single pass
- 100 images per request, with the gain visible in document understanding scores
- Cyber benchmarks place it near the front of open weights models, with 82 percent on CyberGym-E2E-AA
- Weights are coming, so work tuned during the preview transfers straight to self-hosting
- Half price for the first two weeks, putting the trial cost at $0.57 per task
Pricing
Standard pricing is $1.36 per million input tokens and $4.18 per million output tokens, with cached input at $0.14. The first two weeks run at 50 percent off, so $0.68 and $2.09, and OpenRouter's public beta matches this. On Artificial Analysis numbers that works out to roughly $1.13 per intelligence index task at list price and $0.57 during the discount, still above GLM-5.3-Flash at $0.25 and DeepSeek V4.1 Flash at max effort with $0.27. During the preview it is served through the Mistral API and OpenRouter, with self-deployment available once weights are published.
Summary
It helps to read Mistral Large 4 as a release still in progress rather than a finished one. It is a preview today with weights pending, and the date that actually decides where it lands is late October. Once the weights are out it stops being a hosted service from a French vendor and becomes a 1T open weights model anyone can pull, and the open weights ranking on the Cyber Index gets reshuffled along with it.
What is worth noting now is the shape of its capability. Thirty-eight on the intelligence index puts it on the lower edge of the frontier tier, but cost per task is high for that band, still $0.57 during the discount when open models at the same score run a quarter of that. So the value today is not cheap frontier intelligence. It is something else: a frontier model from a European supplier that you will be able to deploy yourself. For teams with residency requirements, or teams who do not want every workload sitting on a US model, that combination has no second option right now.
The multi-image change is the easiest thing to overlook. Going from 8 images per request to 100 moved document understanding up 18 points, which matters far more to filings, tenders and medical records than a leaderboard position does.
If all you want is cheap frontier capability, waiting for the weights costs you nothing. If you want a model you can host yourself after October, the preview is the right time to get the pipeline working.
Version History
- Mistral Large 4 preview (2026-10-06): Mistral released Mistral Large 4 in Research Public Preview: a natively multimodal model with 1T total and 49B active parameters, a 512K context window, up to 256K output tokens and 100 images per request, up from 8. It scores 38 on the Artificial Analysis Intelligence Index, level with GPT-6 Luna at max effort and the highest from outside the US and China, 50 on the Cyber Index with 82 percent on CyberGym-E2E-AA, and 19 percent on GDP.pdf, up 18 points from Mistral Large 3. Standard pricing is $1.36 and $4.18 per million input and output tokens with cached input at $0.14, halved to $0.68 and $2.09 for the first two weeks. It is live on the Mistral API and OpenRouter and entered Agent Arena and Code Arena at 1534 on WebDev in 45th place, with 1T weights planned for the end of October