Mistral releases Mistral Small 3.1

Mistral released Mistral Small 3.1, an update to its 24-billion-parameter open-weight model that added multimodal input — the model could now read images alongside text — and extended its context window to 128,000 tokens. Both the base and instruction-tuned checkpoints were published under the Apache 2.0 licence, permitting unrestricted commercial use and fine-tuning.

Mistral said the model outperformed comparably sized systems, including Google’s Gemma 3 and OpenAI’s GPT-4o Mini, on benchmarks it reported, and described it as capable of running at around 150 tokens per second on suitable hardware — fast enough for tasks like document verification, diagnostics and on-device image processing. The company positioned it as the first open-weight model to match or exceed leading small proprietary models “across all these dimensions,” though as with most vendor-reported comparisons, the benchmark selection was Mistral’s own.

The release continued Mistral’s pattern of pairing frontier-adjacent open releases with a commercial API business: the model was made available on Hugging Face for self-hosting, and simultaneously on Mistral’s own La Plateforme and Google Cloud Vertex AI, with Nvidia NIM and Microsoft Azure AI Foundry listed as forthcoming. At a moment when Meta’s Llama and China’s open-weight labs — DeepSeek, Alibaba’s Qwen — were setting the pace on openly licensed models, Mistral Small 3.1 kept the French lab competitive in the sub-30B weight class most widely deployed on consumer and edge hardware.