Ministral 3
2025-12-02 · Publisher : Mistral AI · Domain : Artificial Intelligence
Mistral AI releases Ministral 3, a range of dense 3B, 8B and 14B parameter models, multimodal, multilingual and under the Apache 2.0 license, in base, instruct and reasoning variants, designed for edge and local use.
What changes
- Three sizes (3B, 8B, 14B), each in base, instruct and reasoning, with image understanding.
- According to the vendor, the best cost/performance ratio among open models, with instruct versions often producing an order of magnitude fewer tokens.
- Reasoning variant of the 14B at 85% on AIME 2025.
- NVIDIA optimization for DGX Spark, RTX PCs and Jetson.
Mistral AI releases Ministral 3, a range of dense 3B, 8B and 14B parameter models, multimodal, multilingual and under the Apache 2.0 license, in base, instruct and reasoning variants, designed for edge and local use. ### What changes - Three sizes (3B, 8B, 14B), each in base, instruct and reasoning, with image understanding. - According to the vendor, the best cost/performance ratio among open models, with instruct versions often producing an order of magnitude fewer tokens. - Reasoning variant of the 14B at 85% on AIME 2025. - NVIDIA optimization for DGX Spark, RTX PCs and Jetson. ### Specifications - 14B: 13.5B language and 0.4B vision, fits in 24 GB of VRAM in FP8; 256k context for the range. - Languages: more than 40 native languages according to the blog post. - License: Apache 2.0. - Availability: Mistral AI Studio, Amazon Bedrock, Azure Foundry, Hugging Face, Modal, IBM watsonx, OpenRouter, Fireworks, Unsloth, Together AI. - API identifiers: `ministral-3b-2512`, `ministral-8b-2512`, `ministral-14b-2512`. ### Pricing | Model | Input (USD / M tokens) | Output (USD / M tokens) | |---|---|---| | Ministral 3 3B | 0.1 | 0.1 | | Ministral 3 8B | 0.15 | 0.15 | | Ministral 3 14B | 0.2 | 0.2 | 8B price confirmed in the blog post; 3B and 14B prices taken from the current documentation. ### Benchmarks Scores published by the vendor (Hugging Face model card): | Benchmark | 14B | 8B | 3B | |---|---|---|---| | AIME 2025 (reasoning) | 85.0% | 78.7% | 72.1% | | GPQA Diamond (reasoning) | 71.2% | 66.8% | 53.4% | | Arena Hard (instruct) | 55.1% | 50.9% | 30.5% | | MATH Maj@1 (instruct) | 90.4% | 87.6% | 83.0% | ### Key takeaways A compact, open family for local and embedded deployments, with a reasoning variant for cases where accuracy comes first.
Links
Tags : LLM, Mistral, Ministral, Edge, Multimodal, Open Source, Apache 2.0