Mistral says its new model is the strongest open-weight AI built outside China. An independent index ranks it 32nd of 44. Both statements are about the same model, Mistral Large 4, which the French lab put into public preview on October 6, 2026. Nicknamed "Le Chonk," it is a mixture-of-experts system with roughly 1 trillion parameters, and the gap between those two verdicts is the most useful thing to understand about it.

The headline claim and the broad ranking disagree

On the broad measure, ML4 is not a leader. DataCamp's launch-day analysis, drawing on vals.ai, puts it at 48.05% on the Vals Index, 32nd of 44 models overall and 9th of 16 among open-weight models. Mistral's own launch post makes a much bigger claim, saying ML4 clearly outperforms every open-weight model developed in the US or Europe and competes with the best open models worldwide.

Put the rankings side by side and a pattern appears. Sixth of 75 on a legal benchmark puts ML4 in roughly the top 8 percent. Fifth of 18 on a cyber index is the top 28 percent. Thirty-second of 44 on the broad index sits about 73 percent of the way down the list. The model is a specialist, and its marketing reads like a generalist's. Mezha noted that the performance claims have not been independently verified.

Cybersecurity is where the evidence holds up

ML4's best independent number is on CyberGym-E2E, a test in which a model must reproduce a real software vulnerability and then patch it. DataCamp, citing Artificial Analysis, reports ML4 scored 82%, first among 18 models. On the wider Artificial Analysis Cyber Index it scored 50%, top five of 18, behind Grok 4.7 and MiMo-V2.6-Pro at 56% each and GPT-6 Luna at 53%.

DataCamp also notes that some closed models score poorly on this test because they refuse many of the tasks, which counts against them. That is Mistral's whole pitch. Defenders often have to reproduce a flaw before they can fix it, and a model that declines is no use at 2 a.m. during an incident. Google took the opposite route with Gemini 4, which, as we reported, launched for vetted cyber defenders only. Mistral wants defenders to run the weights themselves.

The risk is plain. Once weights are public, attackers get the same tool. Pierre Stock, Mistral's vice president of research, told TechCrunch the company will work with trusted partners and governments before release so the weights serve defense. Mistral says cybersecurity leaders, vetted partners and state authorities are red-teaming the model now.

Coding and legal scores are narrower than they sound

Take the legal result first. ML4 scores 15.83% on Harvey's Legal Agent Benchmark, which is sixth of 75 overall and first of 31 open-weight models, according to vals.ai figures reported by DataCamp. A 15.83% score means the model completes roughly one task in six. The ranking is real, but no one should read it as unsupervised legal work.

On coding, Mistral's own chart shows Kimi K3 ahead on the DeepSWE benchmark, 69% to ML4's 61.7%, with the ML4 figure run privately by Artificial Analysis. On Terminal-Bench 4.0, Mistral reports 28.3% while vals.ai's independent run got 22.73%. The five-point difference probably reflects different test setups rather than anyone inflating results, but it shows why a single number from a single harness is not a ranking.

The preview is not the final model

Mistral said at launch that the reinforcement-learning phase of training was not finished, as Quartz reported. So the API model people can test today is not the one whose weights Mistral plans to release on October 27, and every score above may move.

The specifications are also still loose. Mistral's launch post says 1 trillion total and 49 billion active parameters, while its documentation lists 1.05 trillion and 52 billion. Using the documentation figures, only about 4.7 percent of the weights activate per token, which is why the model can be priced like a mid-tier one. DataCamp adds the catch: all the weights still have to sit in memory, so self-hosting needs serious multi-GPU hardware.

The license is the third unknown. One report relayed by DEV Community says ML4 ships under a custom license, not the Apache 2.0 license used for Mistral Large 3, while DataCamp says the terms are unpublished. For a model sold on control, that gap matters.

Europe's "third way" has to survive October 27

French president Emmanuel Macron has called Mistral's approach a third way in AI, per TechCrunch, positioned between closed US models and open Chinese ones. Mistral says it trained ML4 from scratch in its own European data centers on roughly 4,000 Nvidia Grace Blackwell GPUs, which Stock described as two to three times fewer than Chinese rivals use. Nobody has verified that comparison.

The politics are as important as the benchmarks. Euronews noted that the launch came a day after US startup Reflection unveiled its first open-weight model, Beam, as a Western counterweight to Chinese open models. Open weights have become a diplomatic product, as Xi's proposed BRICS open-source AI zone showed last month.

The test comes on October 27, when independent researchers can run the final checkpoint instead of a model still in training. If the cyber lead survives and the license is permissive, Mistral has a defensible niche. If not, the claim to lead outside China will rest on a benchmark ranking its own model does not match.