Oct 29, 2026
Mistral's Mistral Large 4 OSS evaluated across our benchmark suite
We evaluated Mistral’s new open-weights Mistral Large 4 OSS across our benchmark suite.
-
#17 of 30 on the Vals Index (56.42%), #5 of 15 open-weights models, just behind MiMo V2.6 Pro (57.81%).
-
Strongest on Vibe Code Bench (81.56%, #7 of 43 open-weights models) and Harvey’s Legal Agent Benchmark (15.83%, #3 of 16 open-weights models and #8 of 31 overall). It also scores 61.72% on Finance Agent v2.1 (#6 of 19 open-weights models).
-
Weaker on Tax Agent Bench (27.33%, #12 of 29 open-weights models) and Terminal-Bench 4.0 (26.26%, #8 of 22).
Mistral Large 4 OSS has a 1M-token context window. We ran it with temperature 1, top_p 0.95, high reasoning effort and 256k max output tokens.
Congrats to the Mistral team on the release!