Atlas

Models

← All models

Ministral 3 3B

Mistral AI · 2025-12-02 · 3.8B parameters

Ministral 3 3B (`ministral-3b-2512`) is Mistral AI's dense, instruction-post-trained multimodal edge model released on December 2, 2025. Its hosted profile links the FP8 Instruct weights; Mistral publishes the Base and Reasoning versions as separate checkpoints. It accepts text and images, produces text, supports a 262,144-token context through YaRN, and its 3,849,090,048-parameter checkpoint is licensed under Apache 2.0 and can fit in 8 GB of VRAM in FP8.

Benchmark scores

BenchmarkScore
AA-LCR11.7
AA-Omniscience Index-64.1
AIME 2025 (Artificial Analysis avg@10)22.0
Artificial Analysis Agentic Index1.6
Artificial Analysis Coding Index4.8
Artificial Analysis Intelligence Index6.8
Artificial Analysis Omniscience Accuracy7.6
Artificial Analysis Omniscience Hallucination Rate77.5
Artificial Analysis Openness Index38.9
BuseyBench SVG1.3
Capability118.0
CritPt (avg@5)0.0
Design Arena Overall Code Categories1040
FormationEval79.2
GDPval-AA v2285
GDPval-AA v2 (normalized)0.0
Global MMLU Lite56.7
GPQA Diamond (Artificial Analysis avg@5)35.8
Humanity's Last Exam (Text-Only)5.3
IFBench (Artificial Analysis)26.8
LiveCodeBench (Artificial Analysis)24.7
MMMU Pro38.1
ObviousBench25.0
SciCode (Artificial Analysis avg@3)14.4
Terminal-Bench 2.10.0
Terminal-Bench Hard0.0
τ²-bench (Telecom)24.9
𝜏³-Banking4.7
Loading Atlas data…