On October 6, 2026, Mistral AI opened a public preview of its flagship multimodal model, Mistral Large 4 — a sparse Mixture-of-Experts with a total volume of approximately one trillion parameters, of which 49 billion are active, and a context of 1 million tokens. The API is already available in Mistral Studio at a price of $1.36 per million input tokens and $4.18 per million output tokens, and the company promises open weights by the end of October, following a red-teaming phase with cybersecurity, partners, and government agencies. Mistral claims leadership among open-weight models outside China and the best result among all listed models on the security benchmark, but as of the announcement, the evidence base is the vendor's own figures.

image
image

What happened

On October 6, 2026, Mistral AI released the multimodal model Mistral Large 4 into public preview, which has received the unofficial nickname "Le Chonk." The model is built on a Mixture-of-Experts scheme: the total number of parameters is approximately 1 trillion (according to The Decoder — 1.05 trillion), of which 49 billion are active, the context window is 1 million tokens, and a 1.6-billion-parameter vision encoder handles images. The model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own data centers in Europe and supports more than 160 languages, including all official EU languages. The API is open in Mistral Studio (console.mistral.ai) at a price of $1.36 per 1 million input tokens and $4.18 per 1 million output tokens. According to the vendor's figures: 82% on the agentic security benchmark CyberGym-E2E — the best result among all listed models (Claude Opus 5.5 and GPT-6 Astra refuse to perform such tasks due to safety filters), 93% on Cybench, 49.8% on the Coding Agent Index — higher than DeepSeek V4 Pro and Qwen3.8 Max, and 42% versus 41% for GPT-6 Astra in visual grounding on Dense 200. Mistral promises open weights by the end of October 2026.

Context

The main backdrop to the announcement is the long-standing deficit in the open sector: open-weight models have traditionally lagged behind closed flagships, and large-scale training runs in Europe have usually relied on rented compute from hyperscalers. Mistral is betting on a full cycle on its own continent: training was conducted on its own fleet of 3,800 Grace Blackwell GPUs, rather than rented from hyperscalers. The architecture itself is conservative — a sparse Mixture-of-Experts, where about 5% of the parameters are involved in operation, repeating the proven scheme of Mixtral and DeepSeek models; the novelty here is more in scale, infrastructure, and multimodality than in new principles. Another backdrop concerns the comparison methodology itself: leadership in cybersecurity is partly based on the fact that Claude Opus 5.5 and GPT-6 Astra, due to safety filters, do not perform offensive and defensive security tasks, meaning that some competitors are effectively not participating in the measurement.

Why this matters for the industry

For the industry, the announcement is a signal of another round of cost reduction at the frontier class: a result at the level of closed flagships is available via API at a price known from the announcement, and the promised open weights, if the promise is fulfilled, open up independent hosting and fine-tuning for teams — provided that hardware requirements are disclosed. Direct pressure falls on two niches of closed providers: agentic security and financial-legal scenarios, where refusals by Claude and GPT due to safety filters turn into a market deficiency that vendors will have to address in a differentiated way, and the price segment for long-context tasks. If the architecture is confirmed when the weights are opened, the open part of the industry may shift to trillion-scale MoE with multimodality as a new baseline, and the procurement structure will change for companies for which closed models are unavailable or inconvenient — from security tools to document management and geospatial applications with hosting in the EU.

Why this matters for users

The most practical part of the announcement is available today: the API in Mistral Studio at console.mistral.ai allows you to compare Mistral Large 4 with Claude, GPT, and DeepSeek on your real tasks — long documents, diagrams, satellite images, agentic security and financial-legal scenarios, with your own evals instead of public benchmarks. A context of 1 million tokens means that you can hold books, contract collections, or large code repositories in a single request. When open weights are released, there will be an opportunity for local launch or deployment in your own cloud — this is especially interesting for security researchers, teams with sensitive data, and product teams that need hosting in the EU. The model is still in preview: there is no data on latency and stability in the announcement, so it should only be moved to production after your own measurements of quality and speed.

What is still unknown / limitations

All benchmarks claimed in the announcement are vendor-run figures without disclosed methodology: there is no independent reproduction, and verification will only be possible no earlier than the publication of open weights, promised for the end of October 2026. The security comparison is heterogeneous: some competitors do not participate in the measurement due to safety refusals, so leadership in CyberGym-E2E is partly built on the absence of competitors' data. The superiority in visual grounding is a difference of one percentage point without confidence intervals, and a fair interpretation is parity with GPT-6 Astra. For the Coding Agent Index, the number of runs, scaffolding, and possible contamination of benchmarks with training data are unknown. Practical questions remain open: hardware requirements for independent hosting, the exact total size of the model (sources differ: 1 or 1.05 trillion), the timing of the exit from preview, and prices after release.

Sources

Author

Look at AI, editorial team