Atlas

Models

← All models

Mistral Small 4

Mistral AI · 2026-03-16 · 119.4B parameters

Mistral Small 4 is Mistral AI's March 2026 open-weight hybrid model unifying instruction following, configurable reasoning, coding and agentic work, and text-and-image input in a single 128-expert Mixture-of-Experts checkpoint, with four experts routed per token. Mistral publishes rounded sizes of 119B total parameters and 6.5B active per token; the official model card's serving command gives the 256K context limit exactly as 262,144 tokens. The weights are released under Apache 2.0.

Benchmark scores

BenchmarkScore
AA-LCR44.7
AA-Omniscience Index-29.9
ALE-Bench862
Artificial Analysis Agentic Index4.7
Artificial Analysis Coding Index26.6
Artificial Analysis Intelligence Index19.6
Artificial Analysis Omniscience Accuracy22.1
Artificial Analysis Omniscience Hallucination Rate66.8
Artificial Analysis Openness Index38.9
BLXBench67.8
BLXBench Coding category89.3
BLXBench Cost category90.7
BLXBench Debugging category67.9
BLXBench Hallucination category65.0
BLXBench pass rate45.2
BLXBench Reasoning category69.2
BLXBench Refactoring category62.5
BLXBench Security category46.1
BLXBench Speed category95.6
BLXBench UI category77.1
BullshitBench v2 Leaderboard6.0
Capability135.7
CritPt0.3
Eval Connections10.0
Eval Connections — Guess Accuracy22.6
GDPval-AA v2590
GDPval-AA v2 (normalized)4.5
GPQA Diamond76.9
Humanity's Last Exam (Text-Only)9.5
IFBench (Artificial Analysis)48.2
Loading Atlas data…