Mistral AI

Mistral Small 4

Released Mar 16, 2026 · knowledge cutoff unpublished

Status: Unknown
Location: France
Type: LLM
Modality: Multimodal
Context window: 262K
Max output: —
Speed: 172 tok/s · 0.80s to first answer token
Price ($/MTok in / out): $0.15 / $0.60
Cost per task: $0.098
Open weights: Yes

Benchmarks3 of 10 reported

MMLU-Proretired

78%

GPQA Diamond

71.2%

LMArena Elo

1358

SWE-bench Verified

—

SWE-bench Pro

—

Terminal-Bench 2.1

—

AIME (latest)retired

—

Humanity's Last Exam

—

GDPval-AA v2

—

ARC-AGI-2

—

Strengths & weaknesses

Strengths

One small model covers reasoning, vision and coding instead of three separate ones
Fits on a single GPU under a permissive licence
Reasoning can be switched off entirely for latency-sensitive work

Weaknesses

A fast generalist with a real depth ceiling — it won't reason its way out of a hard problem
Merging three previous lines means it is rarely the best at any one of them

256k context and $0.15/$0.60 pricing per Mistral's announcement (earlier trackers listed 128k). MMLU-Pro 78.0 / GPQA 71.2 third-party-reported; LMArena Elo 1358 via BenchLM's Arena tracking. Output speed and time to first answer token are Artificial Analysis' Mistral Small 4 (Reasoning) figures on Mistral's API.

For developers

API model strings

mistralmistral-small-2603

Licence

Apache License 2.0 · Permissive · commercial use permitted

Retirement

No retirement announced

Lineage

No recorded predecessor

News