Mistral Large 4 is a 1-trillion-parameter AI model that Mistral put into public preview on October 6, 2026. You can use it today through Mistral’s API, but the open weights are still only a promise. Here’s what is confirmed, what isn’t, and whether it’s worth your time.
Here are the facts you can scan in ten seconds.
- Preview launched October 6, 2026, on Mistral Studio and the Mistral API only.
- About 1 trillion total parameters, with 52 billion active per request.
- Launch price: $0.68 per million input tokens and $2.09 per million output tokens.
- Open weights are promised “by the end of the month,” with no license published yet.
The short version: the model is real and usable, but the headline “open” part hasn’t shipped.
What Mistral Large 4 actually is
Mistral Large 4 is a mixture-of-experts model. That means it holds about 1.05 trillion parameters in total, but only around 52 billion do the work on each request. This is how huge models stay affordable to run.
It reads both text and images, covers more than 160 languages, and, per Mistral’s docs, handles a 1M-token context window. Mistral says it trained the model from scratch on about 3,800 Nvidia Grace Blackwell GPUs in its European data centers.
One small wrinkle: early coverage said 49 billion active parameters. According to BetaNews, Mistral’s documentation now lists 52 billion, so treat 52 billion as the current number.
What it costs to use now
Mistral is running a launch sale that halves the normal price. Here’s how the two compare.
| Token type | Regular price (per 1M) | Launch price (per 1M) |
|---|---|---|
| Input | $1.36 | $0.68 |
| Output | $4.18 | $2.09 |
Sale prices tend to end without much warning, so don’t build a budget around the lower number. For a rough sense of where this sits, our look at DeepSeek’s API pricing shows how quickly per-token costs shift in this market.
The Mistral Large 4 open weights: what’s promised and what’s missing
Mistral says the weights will arrive by the end of October. That’s the part most developers care about, because it would let you self-host.
The details are thinner than the announcement suggests. The October 27 date comes from The New Stack, not from Mistral, according to Online Tech Tips. The same report says the license may be a custom one rather than the Apache 2.0 license used for Large 3, and Mistral hasn’t published the terms.
That matters because a custom license can limit commercial use. Until it’s out, “open weights” tells you the files will be downloadable, not what you’re allowed to do with them.
Can you run it yourself?
Probably not on a normal PC. The same report estimates around 500GB of memory even with 4-bit quantization, which is far beyond a single RTX 5090.
In practice you’d need a multi-GPU server or a cluster. A maxed-out Mac Studio might manage a heavily compressed version someday, with trade-offs in speed and quality.
If you want a model you can actually run locally, a smaller one is a better fit. Our piece on MAI-Code-1.1-Flash running locally in GitHub Copilot covers a far lighter option.
How good is it? Take the benchmarks with salt
Every benchmark number so far comes from Mistral itself, and none has been independently confirmed. Here’s what the company reports, next to the context that matters.
| Test | Mistral’s claim | Worth knowing |
|---|---|---|
| DeepSWE v1.1 (agentic coding) | 61.7% | GLM-5.3 scores 61%; a live leaderboard puts top rivals near 74% |
| FinWorkBench (finance) | 67% | Matches DeepSeek-V4-Pro |
| Visual grounding (Dense 200) | 42% | GPT-6 Astra scores 41% |
| Blind coding test | 3.74 out of 5 | Second of five, behind Claude Opus 5 |
So it looks competitive, not dominant. Claims that it beats Claude, ChatGPT, or Gemini aren’t backed up, even by Mistral’s own numbers.
Mistral is also pitching the model for cyber defense, and says it’s red-teaming it with security partners who get reduced moderation. That’s a company claim, not an independent finding.
Frequently Asked Questions
Is Mistral Large 4 free to use?
No. It’s a paid API for now, at $0.68 per million input tokens during the launch sale. Free downloads only come if the open weights ship.
When will the Mistral Large 4 weights be released?
Mistral says by the end of October 2026. October 27 has been reported, but that date isn’t from Mistral, so check its announcements.
How many parameters does it have?
About 1 trillion in total (1.05 trillion in the docs), with 52 billion active per request.
Can I run it on my own computer?
Not on a typical one. Even compressed, it needs roughly 500GB of memory, so it’s meant for servers.
My take: try the preview API if you want a European, open-leaning option for coding or document work, especially at the sale price. But wait for the license and independent tests before you commit a project to it, because the “open” label only counts once you can read the terms.
Leave a Reply