Mistral Small 3.1

2025-03-17 · Publisher : Mistral AI · Domain : Artificial Intelligence

Mistral AI unveils Mistral Small 3.1, a 24-billion-parameter model under the Apache 2.0 license that adds image understanding and a 128k-token context to Mistral Small 3.

Mistral AI unveils Mistral Small 3.1, a 24-billion-parameter model under the Apache 2.0 license that adds image understanding and a 128k-token context to Mistral Small 3. ### What changes - Multimodal understanding (text and images) added to the Small model. - Context extended up to 128k tokens. - Improved text performance; announced as ahead of Gemma 3 and GPT-4o Mini in the same category, at 150 tokens/s. - Runs on a single RTX 4090 or a 32 GB Mac (quantized). ### Specifications - 24 billion parameters; 128k context; text and image input, text output. - License: Apache 2.0 (Base and Instruct versions on Hugging Face). - Availability: La Plateforme, Hugging Face, Google Cloud Vertex AI; NVIDIA NIM and Azure AI Foundry announced. - API identifier: `mistral-small-2503` (retired on 11/30/2025 according to the documentation). ### Pricing | Model | Input (USD / M tokens) | Output (USD / M tokens) | |---|---|---| | Mistral Small 3.1 | 0.1 | 0.3 | Pricing confirmed by the Devstral post (“at the same price as Mistral Small 3.1”) and the archived June 2025 pricing page. ### Benchmarks Scores published by the vendor (Hugging Face card, Small 3.1 Instruct): | Benchmark | Score | |---|---| | MMLU | 80.62% | | MMLU-Pro (5-shot CoT) | 66.76% | | GPQA Diamond | 45.96% | | HumanEval | 88.41% | | MATH | 69.30% | | MMMU | 64.00% | | DocVQA | 94.08% | | RULER 128K | 81.20% | ### Key takeaways An open, multimodal base, deployable on-premises, that also serves as the foundation for other models in the range (Devstral, Voxtral).

Tags : LLM, Mistral, Small, Multimodal, Open Source, Apache 2.0