French AI company Mistral released a research preview of Mistral Large 4 on October 6. The company has given it the nickname "le Chonk," roughly "the big one." It's a natively multimodal mixture-of-experts (MoE) model with 1 trillion total parameters, about 49 billion activated per token, and a 512K context window, supporting text and image input with text output.
The preview is currently available through the Mistral API, prioritized for developers, security teams, and government agencies. The company plans to widen access later this month and release open weights by the end of October.
Training and positioning
According to Mistral's official blog, Large 4 was trained from scratch on the company's own data centers in Europe, using roughly 3,800 Nvidia Grace Blackwell GPUs; CNBC reported the training run took about two months. Mistral is positioning the model's strengths around cybersecurity, coding, manufacturing, finance, and multimodal tasks.
Mistral's public claim is that Large 4 is "the most capable open-weight model outside China, by a wide margin." CNBC's report also noted that Mistral acknowledges it still trails the most advanced closed-source models in areas like coding.
The API offers two reasoning modes: none and high. Developer Simon Willison found in his own testing that the high mode actually produced fewer tokens — 2,717 on average versus 3,275 for none — and that outputs from high mode looked better. His overall assessment: Large 4 is roughly half a year behind the frontier.
How third parties score it
Independent benchmarking firm Artificial Analysis gives it an Intelligence Index score of 38. Side by side:
| Model | Intelligence Index |
|---|---|
| DeepSeek V4.1 Flash | 39 |
| Mistral Large 4 | 38 |
| GPT-6 Luna (max) | 38 |
| Mistral Large 3 | 9 |
The 29-point jump between generations is stark — the previous Large 3's score of 9 sits near the bottom of the chart. Artificial Analysis accordingly calls it "the most intelligent model outside the US and China." The firm's Cybersecurity Index gives Large 4 a score of 50, while its document-reasoning benchmark (GDP.pdf) shows a pass rate of just 19% — a clear weak spot.
Doing the math
Preview API pricing is $1.36 per million input tokens and $4.18 per million output tokens, with cached input at $0.14 per million; pricing is 50% off for the first two weeks. Artificial Analysis, converting its own task suite, puts the per-task cost at $1.13, or $0.57 during the discount period.
A rough real-world scenario: feeding in 1 million tokens of a codebase or document set and getting back 200,000 tokens of output costs about $2.20 at list price, or about $1.10 during the discount window. Output is priced at roughly three times the rate of input, so long-output agentic tasks will run more expensive.
The activation ratio is also worth a look. 49 billion activated parameters out of 1 trillion total works out to about 4.9%, in the same range as the day-earlier Reflection Beam release (23 billion activated out of 501 billion total, about 4.6%). Inference cost tracks activated parameters, which is why both companies keep talking up "efficiency."
Deployment is a separate hurdle. At 8-bit precision, just storing the 1-trillion-parameter weights takes roughly 1TB of VRAM, and even 4-bit quantization lands around 500GB. Most teams getting the open weights will likely still need cloud-hosted inference or will wait for community-distilled versions.
Europe's card to play
Mistral closed a €3 billion funding round in September, led by Samsung. Large 4 is the first flagship release funded by that round, and the company keeps emphasizing its in-house compute and European-based training — a sales pitch clearly aimed at European governments and regulated industries.
Mistral hasn't yet published the open-weight license terms or a firm release date. All of the scores so far come only from the company itself and a handful of benchmarking firms; how much of that holds up once the community gets its hands on the weights won't be clear until later this month.
Sources: Mistral official blog, Artificial Analysis, CocoLoop, CNBC, Simon Willison's blog; Intelligence Index and per-task cost figures are as published by Artificial Analysis, API pricing is per Mistral's preview pricing, and VRAM requirements are rough estimates based on parameter count.