France's Mistral claims the most powerful open AI model outside China

About 1 trillion parameters, about 50 billion doing the work

Mistral AI's Mistral Large 4 landed in public preview on 6 October 2026, and the numbers are genuinely large. Mistral's developer documentation puts total parameters at 1.05 trillion, with 52 billion activated per task via a mixture-of-experts design (its launch announcement rounds this to 1 trillion, with 49 billion active), a 1.6 billion-parameter vision encoder, and a 1 million-token context window.^[1] The company nicknamed it Le Chonk, which at least has the virtue of accuracy.

The architecture matters here. Mixture-of-experts routes each query through a subset of parameters rather than firing the full stack, which keeps inference costs tractable at this scale. Running about 50 billion active parameters is expensive but not absurd; running 1.05 trillion would be.

What the benchmarks show, and who ran them

Mistral's own announcement claims ML4 leads open-weight models built outside China across aggregated benchmarks.^[2] Those are company figures, and company benchmark claims deserve the usual scepticism. Mistral says it also had ML4 tested by third-party evaluators, citing vals.ai, which it says found ML4 ahead of GPT-6 Astra on both legal and financial benchmark suites.^[3] Bushletter could not independently verify those figures beyond vals.ai's published results.

On the cybersecurity side, Mistral says ML4 ranks in the top five models globally on the Artificial Analysis Cyber Index, scoring 82 per cent on a vulnerability reproduction-and-patching test and solving 93 per cent of Cybench challenges.^[2] On the public B3 AI Security Benchmark dataset, ML4 resisted 93.3 per cent of indirect prompt-injection attacks, and Mistral said it had seen no higher score among competitors.^[4] Prompt-injection resistance has become a harder requirement as enterprises push AI agents into workflows touching live data.

Pricing and the weight release date

Mistral's pricing documentation lists preview rates at US$0.68 per million input tokens, US$0.07 per million cached-input tokens, and US$2.09 per million output tokens.^[5] Those are introductory figures, so expect movement when the model reaches general availability.

Full model weights are scheduled for public release on 27 October 2026.^[2] That date will determine whether ML4 actually shifts the open-weight frontier: benchmarks on a hosted preview are one thing, weights that developers can self-host, fine-tune and audit are another entirely.

The sovereignty frame and the compute behind it

Mistral trained ML4 from scratch on 3,800 NVIDIA Grace Blackwell GPUs housed in its own European datacentres, framing the release under the slogan "Forged in Europe. Built for AI sovereignty."^[2] The sovereignty pitch lands in Brussels policy circles, where dependence on US hyperscaler compute is a live concern; whether it lands with developers will depend more on the 27 October weight release than on the marketing copy.

Open-weight frontier models have historically carried a qualification: the gap between the best closed models and the best open ones has been wide enough that "competitive" usually means competitive-with-caveats. ML4's claimed benchmark parity with GPT-6 Astra, if it holds under independent scrutiny once weights are available, would close that gap materially, and the next test is three weeks away.