Ministral 3 3B: benchmark results
Mistral's 3B edge model from the Mistral 3 family, released under Apache 2.0 with image understanding and base, instruct and reasoning variants (December 2025). Provider: Mistral. Released 2025-12-02. Access: Open.
Aggregate Rating 1463 ± 16, rank #918 of 1656 rated models, from 149 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| SciImpact - Media | 60.7 | Pairwise prediction accuracy (%, times 100) on media-attenti | 60 |
| HarmActionsEval | 2.13 | SafeActions@1 (imported) | 55.6 |
| 6G-Bench - Distributed Intelligence and Emerging Use Cases | 71.7 | Accuracy (%) on the distributed intelligence and emerging 6G | 51.9 |
| AA-Omniscience Hallucination Rate | 80.19 | Hallucination Rate (%) | 50.3 |
| SciImpact - Chemistry | 61.6 | Pairwise prediction accuracy (%, times 100) on chemistry pai | 50 |
| LinkS2Bench - Zone Counting | 61.9 | Zone Counting: mean relative accuracy (%, MRA over threshold | 47.1 |
| AgentFloor | 42.7 | TCR (self-reported) | 43.8 |
| AA Omniscience - Software Engineering (SWE) - Dart | 16 | Accuracy (%) | 43.7 |
| POLAR-Bench - Utility | 81.46 | Utility score (%): share of the task-relevant attributes the | 42.9 |
| SciImpact - Medicine | 57.7 | Pairwise prediction accuracy (%, times 100) on medicine pair | 40 |
| SciImpact - Patent | 53.6 | Pairwise prediction accuracy (%, times 100) on patent-citati | 40 |
| 6G-Bench - Intent and Policy Reasoning | 81.3 | Accuracy (%) on the intent and policy reasoning tasks (T1, T | 37 |
Interactive version: theaggregate.ai/model?slug=ministral-3-3b · How It Works · Data refreshed daily, snapshot 2026-10-20.