Atlas

Models

← All models

Magistral Small 1.0

Mistral AI · 2025-06-10 · 23.6B parameters

Magistral Small 1.0 is Mistral AI's first small reasoning model, released June 10, 2025 as the open-weight 24B variant alongside enterprise Magistral Medium 1.0. Built on Mistral Small 3.1, it was cold-started with supervised fine-tuning on Magistral Medium traces and then trained with reinforcement learning; its weights are Apache-2.0. Mistral advertised a 128K context but warned that performance might degrade past 40K, while the released checkpoint configuration sets an exact 40,960-position limit; the hosted model retired November 30, 2025 with Mistral Small 4 listed as its replacement.

Benchmark scores

BenchmarkScore
AA-LCR0.0
AA-Omniscience Index-43.4
AIME 202541.3
ARC-AGI-15.0
ARC-AGI-20.0
Artificial Analysis Intelligence Index10.7
Artificial Analysis Omniscience Accuracy18.3
Artificial Analysis Omniscience Hallucination Rate75.5
BRIDGE Medical (chain-of-thought)34.2
BRIDGE Medical (few-shot)49.1
BRIDGE Medical (zero-shot)37.6
Capability130.1
CritPt0.0
Humanity's Last Exam (Text-Only)7.2
IFBench (Artificial Analysis)24.8
LiveCodeBench (Artificial Analysis)51.4
MATH-500 (Artificial Analysis source)96.3
OTIS Mock AIME 2024-202530.0
SciCode (Artificial Analysis)24.1
SnakeBench Average Apples1.9
SnakeBench Best Apples8.0
SnakeBench Rating19.1
SnakeBench Total Apples64.0
SnakeBench Win Rate42.9
Terminal-Bench Hard4.5
τ²-bench (Telecom)26.6
Loading Atlas data…