Mistral Large 4 Benchmarks Trail Chinese Open-Weight Rivals

What Mistral launched

Mistral has introduced Mistral Large 4, a flagship model nicknamed "Le Chonk," with a pitch aimed directly at the geopolitics of open AI. The French company calls it the most capable open model built outside China and says it is the strongest such model "by a substantial margin." 4 It also says that once the weights ship, the model will rank among the top open-weight systems worldwide on aggregate benchmarks. 4

The model was announced on October 6, 2026, and is available as a public API preview. 1 The open-weight release is scheduled for the end of October, and Mistral's Hugging Face page currently lists October 31 as the expected date. 1 No downloadable weights or final license were available at launch. 1 Mistral is first offering the preview to developers, "cybersecurity leaders" and state authorities before a wider release. 4

On paper it is a large system. It uses a granular mixture-of-experts design with about 1.05 trillion total parameters, which Mistral rounds to 1T. Roughly 49 billion parameters are active in the experts, or 52 billion counting embeddings and output layers. 1 It also has a 1.6-billion-parameter vision encoder for native image input. 1 Mistral says it trained the model over two months on 4,000 Nvidia Grace Blackwell GPUs in its own European data centers. 4 The company highlights cyber, coding, manufacturing, finance and multimodal work as strengths, while acknowledging the model still trails the frontier in areas such as coding. 4

What independent numbers show

Third-party measurements describe a large improvement that still leaves the model in the middle of the pack. On the Artificial Analysis Intelligence Index, Large 4 scores 38.4, against 9.3 for Large 3 last December. That is roughly a fourfold jump and well above the 26 median for its price tier. 2 It still sits behind two Chinese open models: Kimi K3 at 43.6 and Qwen 3.8 Max at 45.4. 2

Other figures are less favorable:

• Hallucination: a 41.9% rate on the AA-Omniscience test. 2
• Verbosity: about 200 million tokens to complete the index, compared with an 81 million median. 2
• Speed: 116.1 tokens per second, with roughly 18.7 seconds before the first answer token. 2
• Cost: about $1.13 per index task at list price, or around $0.57 at a promotional rate. 2

The review credits the model with real strengths in cyber tasks without excessive refusals, legal agent workflows and visual grounding. Those areas, however, come from Mistral's own launch charts. 2

The coding claim

The sharpest challenge is to Mistral's coding comparisons. Mistral's launch post says a combined Coding Agent Index score of 49.8% puts Large 4 "ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max." 3 On Terminal-Bench 4.0, one published component, the numbers point the other way. Mistral reports 28.3% for its model, evaluated privately, while Artificial Analysis lists Qwen3.8 Max at 38.9%. 3 OrcaRouter's analysis also notes that Qwen3.8 Max ranks 9th of 138 on Artificial Analysis's coding index, and no comparable independent coding figure exists yet for Large 4. 3

The two independent write-ups also disagree on one point. OrcaRouter lists Large 4's Intelligence Index score as not yet published, while eesel cites 38.4. 23 The likeliest explanation is timing, with the score posted between the two analyses. Either way, the headline result was not available at launch.

Reading the claim carefully

Much of the criticism lands on claims Mistral did not exactly make. Ranking behind Kimi K3 and Qwen 3.8 Max does not disprove "best open model outside China," because both rivals are Chinese. Based on the available data, that narrower claim may well hold.

The weaker claims are the broader ones. Mistral says Large 4 will be among the best open-weight models globally and that it beats Qwen3.8 Max on coding agents. A 7-point deficit on the aggregate index and a 10-point gap on Terminal-Bench make both hard to accept. Building the case on a self-defined combined index, while a component score shows a clear loss, invites skepticism. That is especially true when the open weights and license are not yet public.

Why it matters

Access to the strongest open models has become a point of tension in the U.S.-China technology rivalry, and Chinese developers have held the lead. 4 Mistral is presenting Large 4 as Europe's answer, and the fourfold improvement suggests the company is competitive again. 2

The overall picture is clear, though. Large 4 narrows the gap with China's best open models without closing it. Its verbosity and hallucination rate also raise its practical cost beyond the per-token price. Buyers who mainly want an open model built outside China will find it worth evaluating. Those who simply want the strongest open-weight model, especially for agentic coding, should wait for the October 31 weights and independent testing.