Artificial Analysis benchmarks Mistral Small 4
Artificial Analysis has benchmarked Mistral Small 4, an Apache 2.0 open-weights model with 119B total and 6.5B active parameters, giving builders early cost and speed data on its reasoning and non-reasoning modes.
Independent benchmarking firm Artificial Analysis has published results for Mistral Small 4. The release comes in two variants, Mistral Small 4 (Reasoning) and Mistral Small 4 (Non-reasoning).
According to the listing, the model has 119B total parameters, of which 6.5B are active per token during inference. It accepts text and image input, outputs text, and has a 256k-token context window. Mistral has released it under the Apache 2.0 license, with weights on Hugging Face.
Despite the "Small" name, its total parameter count places it in Artificial Analysis's "Medium" open-weights class (40B–150B parameters). Its results are compared only against other models in that class.
Artificial Analysis reported these figures for its Intelligence Index tasks:
- Cost: the reasoning variant had the lowest cost per task of the two, at $0.0399.
- Speed: the reasoning variant also had the lower time per task, at 1.7 minutes.
- Latency: the non-reasoning variant had the lower time to first answer token, at 1.63 seconds.
The page names the reasoning variant as the release's top model on intelligence. The material available does not include a clear headline index score, results on individual evaluations, per-token API pricing, or how the model ranks against named competitors. Artificial Analysis's capability indexes draw on evaluations including AA-Omniscience, AA-Briefcase, HLE-Diamond, AutomationBench-AA and Terminal-Bench, but the source does not give per-benchmark scores.