Oct 29, 2026 · found Nov 9, 2026, 12:30 PM PT

Mistral Large 4 open weights: ranking claims differ from Artificial Analysis

Verdict: Vendor-reported · Model releases · Benchmarks

The 1T-parameter weights are real and licensed as a modified MIT, but Mistral's coding and cyber numbers are mostly its own and Artificial Analysis ranks the model 11th of 26 open-weight models.

Mistral published weights for Mistral Large 4, a roughly 1T-parameter multimodal mixture-of-experts model, in late October.

Claim: Mistral reports FrontierCode 2.0 of 19.7%, Terminal-Bench 4.0 of 41% and a Coding Agent Index of 48%, and calls it the leading open-weight coding model. Mistral
Evidence: Artificial Analysis lists an Intelligence Index of 26, ranked 11th of 26 open-weight models, at $2.06 per index task. It showed no Coding Agent Index or Cyber Index for the model. Artificial Analysis
Caveats: The Hugging Face post says fifth-best open-weight model, which differs from the AA listing and may reflect a different date or index version. Several figures name no runner, and the parameter count is given as 49B and 52B active. Hugging Face
Why it matters: Hosting it as a service needs a commercial licence. Mistral

Sources: Hugging Face, Mistral Large 4 (primary) · Mistral, Large 4 open weights (primary) · Artificial Analysis, Mistral Large 4 (independent)

Verdict history

Nov 9, 2026, 12:30 PM PT: Vendor-reported. Found late; a Nov 9 digest pointed at it but the Hugging Face post is dated 29 Oct and the model page says weights updated about 10 days ago. Read both Mistral pages and the AA model page. Could not confirm who ran Terminal-Bench 4.0 or FrontierCode runs beyond Mistral's wording.