Mistral Large 4: A trillion parameters and open weights by month-end

Key takeaways

• Mistral Large 4 has 1 trillion total parameters with 49 billion active, using a sparse mixture-of-experts architecture.
• The model was trained on Mistral's own cluster of 3,800 NVIDIA Grace Blackwell GPUs.
• Open weights are promised by the end of October, positioning it as the best open model outside China.
• It scores 38 on Artificial Analysis, a major improvement over Large 3 but still about six months behind the frontier.
• The model is in API preview now with two reasoning modes and multimodal support.

# Mistral Large 4: A trillion parameters and open weights by month-end

What shipped

On October 6, French AI lab Mistral released a preview (https://simonwillison.net/2026/Oct/6/le-chonk/) of Mistral Large 4, a multimodal model with 1 trillion total parameters and 49 billion active parameters. The model is available now through Mistral's API. Mistral has promised to release the open weights version by the end of October.

The model was trained on Mistral's own cluster of 3,800 NVIDIA Grace Blackwell GPUs (https://simonwillison.net/2026/Oct/6/le-chonk/). That infrastructure represents a significant investment and a shift toward owning their training stack rather than relying on cloud providers.

Mistral Large 4 supports two reasoning modes via the API: "none" and "high." The "high" mode appears to produce better results on visual tasks, though testing shows (https://simonwillison.net/2026/Oct/6/le-chonk/) it can use fewer output tokens than the "none" mode in some cases.

How it works

Large 4 is a sparse model. That means it contains 1 trillion parameters total, but only 49 billion are active for any given inference request. This approach, often called mixture-of-experts or MoE, routes each request through a subset of the full model. The result is faster inference and lower cost compared to running all 1 trillion parameters every time.

The model is multimodal. It can handle both text and images as input, though details on training data and modality weighting are not public.

On the Artificial Analysis benchmark (https://simonwillison.net/2026/Oct/6/le-chonk/), Large 4 scores 38, just behind DeepSeek 4.1 Flash (a 552 billion parameter model). That score is a massive jump from Mistral Large 3, released last December, which scored only 9 on the same benchmark. Large 3 also failed basic visual tasks like drawing a pelican on a bicycle. Large 4 handles those tasks competently.

The two-tier reasoning system is simpler than what some other frontier labs offer. "High" reasoning likely involves chain-of-thought prompting or extended generation internally, but Mistral has not published specifics on how the modes differ under the hood.

What it changes for builders

Mistral is positioning (https://www.wired.com/story/mistral-new-model-le-chonk-open-source-china-us-frontier/) Large 4 as the best open-weight model outside China. If the company delivers on the open weights promise, that changes the calculus for teams that need strong performance but want to host models internally or fine-tune them.

Open weights matter most when you cannot use a third-party API. Regulated industries, teams with strict data policies, and companies that need to deeply customize model behavior all benefit. The gap between closed frontier models and the best open models has widened over the past year. A credible open alternative at this scale is useful.

The API preview also matters now. At a score of 38 on Artificial Analysis, Large 4 sits well below the absolute frontier but comfortably ahead of most open models. For tasks that do not require cutting-edge reasoning or vision capabilities, it could be a cost-effective choice. Pricing details are not yet public.

Our read is that Mistral is re-entering the conversation after a quiet period. Large 3 was widely seen as a disappointment. Large 4 will not beat GPT-6, Claude Opus 5, or Gemini 3, but it does not need to. It needs to be good enough that developers consider it, and good enough that hosting it on your own infrastructure makes sense. Simon Willison notes (https://simonwillison.net/2026/Oct/6/le-chonk/) it is "maybe about 6 months behind the frontier." That is a competitive position for an open model.

Gotchas and limits

The model is in preview. Performance, pricing, and availability could all change before the open weights release.

Reasoning modes are limited to two levels. If your application needs fine control over compute-per-request or reasoning depth, you will need to handle that in your own prompt design or retry logic.

We do not know the license terms yet. "Open weights" does not always mean unrestricted commercial use. Mistral's past licenses have varied. Confirm terms before planning a deployment.

The 49 billion active parameter count is lower than some competing sparse models. DeepSeek 4.1 Flash, for example, has 552 billion parameters total. The architecture matters as much as the raw parameter count, but if you are comparing models on paper, keep the sparsity ratio in mind.

Vision performance is improved but not frontier-level. If your use case is vision-heavy, test carefully against task-specific benchmarks or your own evals before committing.

How we would use it

We would wait for the open weights release and the license before making any commitments. Once available, we would consider Large 4 for clients who need multimodal capabilities, have data residency requirements, and do not need the absolute best model on every task.

The API preview is worth testing now if you are already evaluating models in the 30 to 40 Artificial Analysis range. Compare cost and latency against DeepSeek, Llama 4, and other mid-tier offerings.

For fine-tuning or hosting, the 49 billion active parameter count is manageable on modern GPU infrastructure. A team with access to A100s or H100s could run inference without exotic setups. If your workload justifies the engineering effort, this could be a strong candidate for an internally hosted model.

We would not use this for tasks that require frontier reasoning or vision. But for summarization, classification, structured extraction, and straightforward multimodal tasks, it could be a practical choice once the weights are public and the license is clear.

Mistral is back in the race. Whether they stay there depends on what they ship next and how quickly.