Atlas

Models

← All models

Ministral 3 3B Instruct 2512

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

Ministral 3 3B Instruct 2512 is Mistral AI's December 2025 instruction-post-trained, non-reasoning checkpoint for edge and local deployment. The open-weight Apache-2.0 model combines a dense 3.4B-class decoder with a 410M vision encoder, accepts text and image input, produces text, and supports a 262,144-token context; this repository is the FP8 checkpoint, while Base and Reasoning are separate releases. Mistral's API serves it as ministral-3b-2512 (model card: https://docs.mistral.ai/models/model-cards/ministral-3-3b-25-12).

Benchmark scores

BenchmarkScore
AA-Briefcase Elo16.0
AA-Briefcase v1.1 Elo0.0
AA-LCR11.7
AA-LCR v1.117.0
AA-Omniscience Index-64.0
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 Intelligence Index v4.34.8
Artificial Analysis Intelligence Index v4.3.24.8
Artificial Analysis Omniscience Accuracy9.0
Artificial Analysis Omniscience Hallucination Rate80.2
Artificial Analysis Openness Index38.9
Artificial Analysis Progress Index v1150
AutomationBench-AA 1.0.60.7
BuseyBench SVG1.3
Capability115.9
CritPt (avg@5)0.0
Design Arena Overall Code Categories1008
DTBench51.7
FormationEval79.2
GDP.pdf (Artificial Analysis, All-pass)0.0
GDP.pdf (Artificial Analysis, Mean Criteria)16.5
GDPval-AA v2234
GDPval-AA v2 (normalized)0.0
GDPval-AA v2.137.1
Global MMLU Lite56.7
GPQA Diamond (Artificial Analysis avg@5)35.8
Humanity's Last Exam (Text-Only)5.4
Loading Atlas data…