Model Details
MiniStral 3 3B Instruct 2512
Quick answer: MiniStral 3 3B Instruct 2512 is Mistral AI's December 2025 compact 3B model scoring 83.0% MATH, 54.8% LiveCodeBench, and 30.5% Arena Hard. MRL v0.1.
At a Glance
Where MiniStral 3 3B leads
- 83.0% MATH — strong competition math for 3B
- 54.8% LiveCodeBench — competitive real-world coding
- Compact 3B: low inference cost
Where it lags
- 30.5% Arena Hard — moderate instruction quality
- MRL v0.1: not Apache 2.0
Best for: Compact Mistral deployments at 3B scale; math and coding workloads.
What MiniStral 3 3B Instruct 2512 Is
MiniStral 3 3B Instruct 2512 is part of Mistral's December 2025 MiniStral 3 release — compact instruction models with MRL v0.1. The 83.0% MATH shows strong mathematical ability at 3B scale.
Specifications
| Field | Value |
|---|---|
| Organization | Mistral AI |
| License | MRL v0.1 |
| HuggingFace | mistralai/MiniStral-3-3B-Instruct-2512 |
| Release date | December 2025 |
| Parameters | 3B |
| Modality | Text only |
Pricing
Available via Mistral API and open weights under MRL v0.1.
Public Benchmark Scores
| Benchmark | Score | Source | Date |
|---|---|---|---|
| MATH | 83.0% | Benchgen evaluation | 2025-12 |
| LiveCodeBench | 54.8% | Benchgen evaluation | 2025-12 |
| Arena Hard | 30.5% | Benchgen evaluation | 2025-12 |
MiniStral 3 3B vs Alternatives
| Model | MATH | LiveCodeBench | Arena Hard | License |
|---|---|---|---|---|
| MiniStral 3 3B 2512 | 83.0% | 54.8% | 30.5% | MRL v0.1 |
| MiniStral 3 14B 2512 | 90.4% | 64.6% | 55.1% | MRL v0.1 |
| Mistral Small 3 24B | 70.6% | — | 87.6% | Apache 2.0 |
Related Models & Benchmarks
- Compare: MiniStral 3 14B 2512 · Mistral Small 3 24B
- Benchmarks: MATH · LiveCodeBench · Arena Hard
Frequently Asked Questions
What is MiniStral 3 3B Instruct 2512?
Mistral AI's December 2025 compact 3B model scoring 83.0% MATH, 54.8% LiveCodeBench. MRL v0.1.Specs from Mistral AI's MiniStral 3 3B Instruct 2512 release (December 2025) and Benchgen evaluations. Last updated 2026-07-24.
Benchmark Leaderboards
This model isn’t on any benchmark leaderboard yet.