Overview
Mistral Large 4 is a frontier model Mistral released on 6 October 2026: 1T total parameters with 49B active in a sparse design, natively multimodal, 512K context and up to 256K output. It launched as a Research Public Preview on Mistral's own API and OpenRouter, and on 29 October the weights were opened, published on Hugging Face as Mistral-Large-4-1T-A52B under the codename le Chonk. The technical report followed on 3 November, where the team describes the open-weight build as more efficient and more compact than the preview from three weeks earlier.
Weight shows up in the scores. Artificial Analysis gives it an intelligence index of 38, level with GPT-6 Luna at max and just under DeepSeek V4.1 Flash at max with 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 Code Arena's WebDev board at 1534 points and 45th place, 304 points above its own Mistral Large 3.
Key Features
- 1T parameters, 49B active: A sparse design with a trillion total parameters and 49 billion active per token, trading large capacity for per-token compute. That is what lets it sit at the edge of the top tier on the intelligence index while staying serviceable on cost.
- Natively multimodal, 100 images per request: Text and image in, text out. The API accepts 100 images per request where earlier Mistral models took 8, and the gain shows in document understanding: 19% on GDP.pdf, 18 points above Mistral Large 3.
- 512K context with 256K output: A 512K token window and up to 256K tokens out in one call, enough for long repositories, long documents and long conversations.
- Cybersecurity in the top tier: A Cyber Index score of 50 from Artificial Analysis, level with GLM-5.3-Flash. With weights open it enters the top three open-weight models on that board. Its strongest result is 82% on CyberGym-E2E-AA, above MiMo-V2.6-Pro at 79% and GPT-6 Luna max at 78%.
- Verifiable on coding and agent boards: It appears in Arena's Agent Arena and Code Arena across WebDev, Text and Vision. WebDev puts it at 1534 points and 45th, a large jump from Mistral Large 3 at 130th, and above Mistral Medium 3.5.
- Open weights you can deploy: The 1T weights released on 29 October sit on Hugging Face as Mistral-Large-4-1T-A52B, codenamed le Chonk. For scale, Mistral trained it from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centres, which is the hardware class a self-hosted deployment has to plan for.
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
- An intelligence index of 38, the highest score from outside the United States and China.
- 512K context with 256K output, so long tasks fit in one pass.
- 100 images per request, with the multimodal document gain reflected in the scores.
- Cybersecurity benchmarks put it at the front of open-weight models, with 82% on CyberGym-E2E-AA.
- Weights are open on Hugging Face, so workflows tuned during the preview move straight to self-hosting.
- The technical report is public, so its limits can be recomputed on your own terms.
Pricing
Standard pricing is 1.36 USD per million input tokens and 4.18 USD per million output tokens, with cached input at 0.14 USD; the first two weeks ran at half price, 0.68 and 2.09 USD, matching the OpenRouter beta. On Artificial Analysis accounting a standard-price intelligence task costs about 1.13 USD, above GLM-5.3-Flash at 0.25 USD and DeepSeek V4.1 Flash at max with 0.27 USD, so it does not sit in the cheap bracket. Hosted access runs through the Mistral API and OpenRouter; with weights open on Hugging Face, self-hosting leaves hardware and operations to pay for.
Summary
Mistral Large 4 settled into its position in late October. Weights opened on 29 October and the technical report followed on 3 November, turning a French company's hosted service into an open-weight model anyone can pull down and run, with the Cyber Index board reshuffling to match.
The shape of its capability is worth noting. An intelligence index of 38 puts it at the lower edge of the top tier, and cost per task in that tier runs high, about 1.13 USD at standard pricing where open models at the same score reach a quarter of that. Its value is therefore not savings but a combination: a frontier model from a European supplier that you can deploy yourself without dropping a tier. For teams with residency requirements or a reluctance to put every workload on American models, that combination has no second option today.
The image handling is the least noticed improvement. Going from 8 images per request to 100 lifted document understanding 18 points, a change far more sensitive to work like financial reports, tenders and medical records than any leaderboard position.
The cost is the deployment bar. A trillion-parameter model does not fit on one machine, so self-hosting means a cluster or on-demand cloud capacity; validating first is easier through the Mistral API or OpenRouter at 1.36 USD input and 4.18 USD output per million tokens.
Version History
- Technical report published, open-weight build tighter than the preview (2026-11-03): Mistral published the technical report for Mistral Large 4, covering how the 1T-parameter, 49B-active model was built, readable on the Hugging Face repo and alphaXiv. The team describes the open-weight release as frontier-capable on coding agents and cybersecurity, and more efficient and more compact than the preview from three weeks earlier. Weights are on Hugging Face, and the model is also reachable through Mistral Studio and the API.
- Open weights released: 1T total parameters, 49B active (2026-10-29): Mistral released the weights of Mistral Large 4, codenamed le Chonk, on October 29: a natively multimodal model with 1 trillion total and 49 billion active parameters, trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centre. Mistral reports 47 percent token savings for coding agents, with weights on Hugging Face and access through Mistral Studio and the API.
- Responses API enters public preview (2026-10-28): Mistral released the Responses API in public preview, built for agent development on its platform. Each response returns the answer, the reasoning trace and tool calls as separate items, which is easier to step through than Chat Completions, and it works with every chat model Mistral hosts, including Large 4.
- Mistral Large 4 enters the Agent Arena top fifteen labs (2026-10-09): Arena announced Mistral Large 4 has entered the top fifteen labs on Agent Arena. Based on more than 5,000 real agent sessions, the preview posts a net improvement score of -6.6% and ranks 43rd overall, 11 places above Mistral Medium 3.5 (-12.60%).
- 1534 on Code Arena WebDev, ranked 45th (2026-10-07): Arena announced that Mistral Large 4 entered the Code Arena WebDev leaderboard at 1534 for 45th place, 304 points above Mistral Large 3 in 130th and 271 points above Mistral Medium 3.5
- 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