Mistral Large 4 preview: 1T parameters, 49B active, open weights due

On October 6, 2026, Mistral released a preview of Mistral Large 4. Simon Willison covered it in Introducing Mistral Large 4: Le chonk, and his write-up is the basis for the details below. The headline numbers are 1 trillion total parameters and 49 billion active parameters. Mistral trained it on its own cluster of 3,800 NVIDIA Grace Blackwell GPUs.

What 1T total and 49B active means

The gap between total and active parameters means each token is processed by only a fraction of the weights. Per-token compute therefore scales with the 49B active figure, not the full trillion. Memory is different: you still have to hold all the weights somewhere. If you plan to self-host, plan around the total size. That is a very different footprint from a dense model with the same compute cost.

Preview now, open weights later

The preview is available through Mistral's API. Mistral promises to release the open-weights model at the "end of this month". For teams that need on-prem or air-gapped inference, such as regulated fintech and trading shops, that date matters more than the preview. Treat it as a promise until the weights actually land. We could not confirm the license, context window or pricing from the sources we checked, so verify them on Mistral's announcement before you plan around them.

Only two reasoning levels

Through the Mistral API the model supports just two reasoning levels, none and high. Willison's pelican test shows why that is worth benchmarking yourself. The high setting gave the better drawing but used fewer output tokens (2,717) than none (3,275). That is one anecdote, not a benchmark, but it suggests the reasoning toggle will not always map to "more tokens, more cost".

Where it lands

Willison reports that the model scores 38 on the Artificial Analysis index. That puts it just behind DeepSeek 4.1 Flash and well above Mistral Large 3, which scored 9. He describes it as roughly six months behind the frontier. Those are third-party framing and index scores, not guarantees for your workload.

What builders should do

• Run it on your own evals, such as tool-calling reliability, structured output and long-context retrieval, rather than trusting an index score.
• Test both reasoning levels and measure real token usage and latency per task.
• If you need self-hosting, wait for the weights and check the license terms before committing.
• Keep your model layer swappable. A preview can change before general availability.

For agentic and trading-adjacent systems, a stronger open-weights model is useful mainly because it lets you keep sensitive data in your own environment. It does not remove the need for guardrails around what the model is allowed to do.