Mistral Large 4 Le Chonk: Open-Weight Claim Outpaces Reality

What Mistral launched

French AI lab Mistral has introduced Mistral Large 4, nicknamed "Le Chonk." It is a mixture-of-experts model with roughly 1 trillion total parameters (1.05 trillion on the model page) and 49 billion active per token.3 It went live as a public API preview on October 6, 2026. It accepts text and image input through a 1.6-billion-parameter vision encoder, offers a 1-million-token context window, and was trained on data covering more than 160 languages.34

Mistral's framing is ambitious. The company calls Le Chonk the strongest open-weight model built outside China and says it comes "very, very close" to some proprietary systems.5 It also stresses that the model was trained from scratch in Europe.52 That is a pointed contrast with Chinese labs, which the US government has accused of leaning on distillation from larger American models to narrow the gap.5 Mistral says the model is tuned for coding and cyberdefense, along with specialist work in manufacturing, finance, and electrical engineering.5

The open-weight promise, pending

For now, "open-weight" describes a plan, not a product. The weights are not downloadable. Mistral says they will arrive by the end of October, and reporters were reportedly given October 27 as the target date.4

The company says the delay exists so it can red-team the model with cybersecurity partners and state authorities before public release.2 Given how well the model performs on security tasks, that caution is plausible. Still, the license question is unresolved. The model page describes the license only as "Open" and does not state the terms.34

This matters because license terms decide whether an open-weight model is usable commercially. One review concludes that enterprises should hold off on any self-hosting commitment until both the weights and the license are actually published.1 Wired's description of a model that "can be used and customized by anyone" reflects Mistral's stated intent. It is not yet something anyone can verify.5

Benchmarks: strong in security, mixed elsewhere

The independent numbers paint a more modest picture than the launch messaging. On the Artificial Analysis Intelligence Index, the preview scored 38.24 How that score reads depends on the comparison set.

• One account says it placed 25th of the 25 models tracked.2
• Another compares it with other open models. On that view it sits well ahead of US open models such as Nemotron 3 Ultra (23) but 6 to 8 points behind Chinese rivals GLM-5.3, Kimi K3, and MiMo-V2.6-Pro.4

The two readings are not contradictory. Le Chonk can be the best non-Chinese open-weight model and still trail the overall field.

On coding, the picture is similar. Its 61.7% on DeepSWE v1.1 falls below published scores from Kimi K3 and DeepSeek V4.1 Flash.3 Results in finance, mathematics, document analysis, and long-context tasks are described as uneven. One review singles out a single strong legal-drafting result.1

Cybersecurity is where the model clearly stands out:

• It tied for first on the Artificial Analysis Cyber Index with GLM 5.3 Flash.2
• It ranked in the top five globally on CyberGym.2
• It showed exceptional performance in finding and patching vulnerabilities.1

That strength fits Mistral's stated focus. It also fits the company's explanation for running security vetting before releasing the weights.

Price and hardware

The list price for API access is $1.36 per million input tokens and $4.18 per million output tokens. That is cheap compared with closed frontier models.24 Mistral's own model page currently shows half those rates, $0.68 and $2.09, with no explanation of why or for how long.4

Self-hosting will be difficult even after the weights ship. The weights alone take about 525 GB at INT4, 1.05 TB at FP8, or 2.1 TB at BF16. That makes an eight-GPU B200 server the most realistic baseline.1

Reading the launch

Mistral's headline claim is narrow, and on the available evidence it is defensible. The company appears to lead among open-weight efforts from the US and Europe.4 But the claim depends on carefully chosen boundaries. Measured against the full field, including Chinese open models and closed frontier systems, Le Chonk is a capable mid-tier contender rather than a leader.124

The more substantial story is political. Wired presents the release as Mistral showing it is still competing as access to top-end AI becomes contested.5 A from-scratch European model with real security capability has value for sovereignty-minded buyers, even if it does not top the overall leaderboards.

For developers and security teams, the practical advice is the one reviewers converge on: pilot through the API now and defer infrastructure decisions.1 Until downloadable weights and a named license exist, Le Chonk is effectively a promising, low-cost hosted model. Whether it earns the "open" label depends on what Mistral publishes by the end of October.