Ministral 3B / 8B (24.10)

// Overview

Ministral 3B and Ministral 8B, collectively nicknamed "les Ministraux," are two small language models that Mistral AI released on October 9, 2024 (and announced publicly on October 16). They were Mistral's first models purpose-built for the sub-10-billion-parameter edge tier — small enough to run on phones, laptops, tablets and IoT hardware — and were positioned for local, privacy-first uses such as on-device translation, offline smart assistants, local analytics and autonomous robotics.

Both models ship with a 128,000-token context window (vLLM was initially capped at 32k because the interleaved sliding-window attention kernels were not yet implemented there), native function calling, and training on a large proportion of multilingual and code data across ten languages: English, French, German, Spanish, Italian, Portuguese, Chinese, Japanese, Russian and Korean. The Ministral-8B-Instruct-2410 weights are openly downloadable on Hugging Face under the Mistral Research License, while Ministral 3B was offered only through the API under a commercial license.

The Ministraux were retired relatively quickly: Mistral marked both ministral-3b-2410 and ministral-8b-2410 as deprecated on December 2, 2025 and removed them from la Plateforme at the end of December 2025, pointing users to the newer Ministral 3 (25.12) generation. The open 8B weights, however, live on for self-hosted research and remain a popular small base for fine-tuning.

// Benchmarks

1. MMLU (Ministral 3B, base)60.9%
2. MMLU (Ministral 8B, base)65%
3. HumanEval pass@1 (Ministral 8B Instruct)76.8%
4. Math maj@1 (Ministral 8B Instruct)54.5%
5. MBPP pass@1 (Ministral 8B Instruct)70%
6. Arena Hard (Ministral 8B Instruct)70.9%
7. Wild Bench (Ministral 8B Instruct)41.3%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

Launch pricing on Mistral's la Plateforme; both models were retired from the API at the end of December 2025.

Pricing source ↗

// Strengths

• Strong accuracy for its size — Ministral 3B scored 60.9 on MMLU and Ministral 8B 65.0, beating Gemma 2 2B, Llama 3.2 3B and edging Llama 3.1 8B on knowledge benchmarks
• 128k-token context with interleaved sliding-window attention for memory-efficient long-context inference
• Native function calling, making the 8B a capable agent backbone for input parsing, task routing and API calls
• Designed to run locally on edge hardware for privacy-first, offline-capable inference
• Open Ministral-8B-Instruct-2410 weights on Hugging Face for research and fine-tuning
• Very low API pricing during its lifetime — $0.04/M (3B) and $0.10/M (8B)

// Best for

• On-device and edge inference on phones, laptops and IoT devices
• Privacy-first, offline-capable assistants and on-device translation
• Local analytics and lightweight text processing without cloud round-trips
• Agentic workflows using function calling for task routing and API orchestration
• Cost-sensitive, high-volume text generation where a small model suffices
• A small open base (8B) for research and domain fine-tuning

// How to access

// Ministral — every version

The full lineage of the Ministral line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.