On October 6, a Paris lab did something no European company had done before: it released a preview of a 1-trillion-parameter AI model, and promised to give the weights away by the end of October. The model is Mistral Large 4, nicknamed “Le Chonk,” and whether you care about open models, European tech sovereignty, or just getting more intelligence per dollar, it deserves your attention.
This is not another press-release rewrite. Here is what Large 4 actually is, what independent benchmarks say about it, and how its pricing reshapes the bill for anyone running models at scale.
A trillion parameters, priced like a 50-billion one
Mistral Large 4 is a natively multimodal Mixture-of-Experts model with roughly 1.05 trillion total parameters. Only about 52 billion “active” parameters engage for any given token, according to Mistral’s model card. (Early reports said 49 billion; use the 52 billion figure.) Think of it like a hospital with a thousand specialists on staff: for any single patient, you only consult the handful relevant to the case. That is the MoE trick, and it gives frontier-scale knowledge with inference costs closer to a mid-size dense model.
The rest of the spec sheet is equally serious: a 1.6-billion-parameter vision encoder bolted onto the MoE core for native image input, a one-million-token context window (though some API providers cap it around 524K), and fluency Mistral claims spans more than 160 languages, including every official EU language, across Latin and non-Latin scripts.
Trained in Europe, on its own machines
The part of the story Mistral wants you to remember is where the training happened. The company built Large 4 from scratch in about two months on roughly 3,800 Nvidia Grace Blackwell GPUs running in its own European datacenters. It did not rent a slice of someone else’s cloud, a choice that reflects a bet on control and aligns with Europe’s push for technological sovereignty. French President Emmanuel Macron has spoken of a “third way” in AI. Large 4 embodies one version of that idea.
The ambition is expensive, and Mistral is funded for it. The company says Large 4 is the first milestone on the roadmap funded by its €3 billion Series D, which it calls the largest equity round ever by a European technology company, with Samsung leading. CEO Arthur Mensch told a conference in Abu Dhabi that the model beats Chinese open models on some aspects, “including cyber,” Reuters reported.
What it costs against the closed frontier
Here is the arithmetic that matters for anyone with an API bill. Large 4 lists at $1.36 per million input tokens and $4.18 per million output tokens, with cached input at $0.14. A 50 percent launch discount currently brings that to $0.68, $2.09, and $0.07, though Mistral has not announced when the sale ends.
| Model | Input / 1M tokens | Output / 1M tokens |
|---|---|---|
| Mistral Large 4 (launch price) | $0.68 | $2.09 |
| Mistral Large 4 (list price) | $1.36 | $4.18 |
| GPT-5.5 | $5.00 | $30.00 |
| Gemini 4 Argon (until Nov 30) | $2.00 | $10.00 |
| Meta Muse Wonder | $4.00 | undisclosed |
Against GPT-5.5’s $5 and $30 pricing, Large 4 is roughly ten times cheaper on a blended basis. It undercuts even Gemini 4 Argon, whose $2/$10 pricing doubles after November 30, and Meta’s Muse Wonder at $4 per million input tokens. For a company running thousands of API calls a day through an internal tool, that gap matters more than a few points on a reasoning benchmark ever will.
Where it wins, and where it does not
Start with the honest part. On Artificial Analysis’s Intelligence Index, Large 4 scores 38, a fourfold jump from Mistral Large 3’s 9, but it trails well behind frontier closed models like Claude Opus 5.5 and GPT-6 Astra on hard reasoning and agentic tests such as Terminal-Bench and Humanity’s Last Exam. An independent cross-check by VentureBeat put it at 62 percent on DeepSWE against GLM-5.3’s reported 69 percent, confirming that Chinese open-weight models still hold a coding benchmark edge at launch.
Where Mistral claims leadership, the numbers are more interesting. The company says Large 4 beats GPT-6 Astra on the Finance Agent v2 test (54.7 vs 53.5), scores best among open models on SciCode-Verified, and leads its AutomationBench workflow-automation benchmark at around a fifth of the per-task cost of rival models. On cybersecurity, its pitch wedge, it reports 93 percent on Cybench and 82 percent on a vulnerability reproduction and patching test, noting that several leading closed models score near zero on the same test because they refuse to perform the task.
Simon Willison, who tested the model on launch day, flagged a real limitation: the API currently exposes only two reasoning levels, “none” and “high,” with nothing in between. Mistral is positioning the model for programming, cybersecurity, manufacturing, finance, and corporate workflows rather than claiming it beats the closed frontier on the hardest tasks.
What “open” actually means right now
The weights are not out yet. Developers can use Large 4 today through Mistral’s API on Mistral Studio, and a preview version with reduced safety barriers is going to select cybersecurity experts and government authorities first. The full weights are promised by the end of October, with Reuters reporting October 27, and will publish under a custom Mistral license rather than a standard open-source one.
That timing caveat is the counter-case to the whole launch: if the weights slip, the license is restrictive, or the final build underperforms the preview, Large 4 is a decent mid-priced API in a crowded tier rather than an open-weight event. Mistral says a final version will ship with the weights, so the downloadable checkpoint may differ from today’s preview.
Still, the signal is clear. The open-weight frontier in 2026 has been overwhelmingly a China-versus-everyone-else race, and Mistral is positioning itself as the one credible non-Chinese entrant still competing at the top of that table. If the weights land on schedule, developers who could never afford GPT-5.5 at scale will have a trillion-parameter model they can run, fine-tune, and host themselves. The bill is where open models win, and Le Chonk just made the bill a lot smaller.
