Mistral Large 4: Specs, Pricing, Open Weights and How to Use It

Mistral Large 4 is Mistral AI's new flagship, a natively multimodal mixture-of-experts model with about 1 trillion total parameters and 52 billion active per token. It launched on 6 October 2026 as a public preview in Mistral Studio and the API, with a 1M-token context window and support for 160+ languages. Mistral says the open weights arrive at the end of October 2026. List pricing is $1.36 per million input tokens and $4.18 per million output tokens, and the docs currently show a half-price sale.

Key facts

What Mistral Large 4 is

Mistral's announcement calls ML4 (nicknamed "le Chonk") a "1 trillion-parameter natively multimodal model with 52 billion active parameters", trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters. The preview is served on the same infrastructure, with a European deployment Mistral operates end to end, plus other regions.

Three things set it apart from Mistral's earlier Large models:

• Scale. At over a trillion total parameters it's in the same weight class as the biggest open-weight releases, such as Muse Spark Open (816B).
• One model for chat and reasoning. Mistral describes it as a hybrid instruct-and-reasoning model; the docs show a reasoning mode.
• Image understanding. Mistral highlights charts, documents, natural images, technical drawings and satellite imagery.

If "52B active" is unfamiliar, our mixture of experts entry explains why a trillion-parameter model can run with the compute cost of a much smaller one.

Benchmarks Mistral published

All of these are Mistral-reported. Independent results will come once the weights are out.

Mistral mostly compares against other open-weight models: it claims a lead over DeepSeek V4 Pro 0813 and Qwen3.8 Max on its Coding Agent Index, and over Kimi K3 and DeepSeek V4 Pro on AutomationBench. Against closed frontier models the picture is mixed. Its own human coding eval puts Claude Opus 5 ahead, and a Terminal-Bench 4 score of 28.3% is well below the 57.9% OpenAI reports for GPT-6 Astra and the 57.4% Google reports for Gemini 4 Argon on Terminal-Bench 4.0. The honest reading: a strong open-weight model, with cyber and agentic workflows as its standout areas, not a closed-model killer.

How to use Mistral Large 4

During the preview: Mistral Studio and the API

1. Sign in to Mistral Studio (Mistral's console) and create an API key.
2. Find Mistral Large 4 in the model list and copy its exact model name. The docs page for it lives at mistral-large-4-0; confirm the identifier in Studio before you hard-code it.
3. Call it through /v1/chat/completions or the Agents and Conversations endpoints.

The docs list support for structured outputs, function calling, document Q&A, prefix, batching (/v1/batch), agents and built-in tools. That covers most agent setups; for the concepts, see function calling and how to build an AI agent.

After the weights land: self-hosting

Mistral says weights "drop end of this month" (October 2026), and the announcement mentions private-cloud and on-premise deployment for security-sensitive work. At 1.05T parameters, self-hosting is a multi-GPU datacenter job even when quantised; it's not a laptop or single-workstation model. If you want something you can run at home, see run an LLM locally.

Before committing, wait for the licence. Mistral Large 3 shipped under Apache 2.0, but Mistral hasn't said what Large 4 will use, and that determines whether you can fine-tune and resell it freely.

Pricing (as of October 2026)

Mistral doesn't give an end date for the sale price, and its general pricing page doesn't list Large 4 yet. Budget at list price and treat the discount as a bonus.

Limits and open questions

• Preview status. Behaviour and pricing can change before general availability.
• Licence unknown. "Open-weight" without a named licence isn't enough to plan a product on.
• Parameter count wording. The announcement says 1T; the docs say 1.05T plus a 1.6B vision encoder. The docs are the more precise figure.
• Max output not published.
• Benchmarks are self-reported and mostly against other open-weight models.

Who should look at it

European teams that need data kept in the EU, organisations planning an on-prem frontier model, and anyone running cybersecurity or multilingual workloads should put Mistral Large 4 on their evaluation list now, through the API, and revisit when the weights and licence are out.