Mistral Small 3.1
Mistral AI
Mistral just dropped Small 3.1, an open model that reads images and beats Gemma 3 and GPT-4o Mini in its size class. It's free under Apache 2.0 and runs on a single gaming GPU.
Based on reporting by Mistral AI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Mistral AI has a habit of releasing models that punch above their weight, and Small 3.1 keeps that streak going. It builds on last month's Small 3 but adds image understanding, a 128k-token context window, and inference speeds around 150 tokens per second. The company claims it now outperforms Google's Gemma 3 and OpenAI's GPT-4o Mini across text, multimodal, multilingual, and long-context benchmarks — not just matching proprietary rivals, but beating them, while staying fully open.
That openness is the real headline. Mistral is shipping both the pretrained base model and the instruct-tuned version under Apache 2.0, meaning anyone can download, modify, or commercialize it without asking permission. That matters because Small 3.1 is genuinely lightweight: it fits on a single RTX 4090 or a 32GB-RAM Mac, which puts a capable multimodal model within reach of hobbyists and small dev teams, not just cloud-scale labs.
Mistral is pitching it for the usual enterprise wishlist — conversational assistants, function calling for agentic workflows, fine-tuning for specialized fields like legal or medical work — but the more interesting signal is what the community has already done with the previous version. Nous Research built its DeepHermes 24B reasoning model on top of Small 3, and Mistral seems keen to keep that pipeline going by releasing raw base weights alongside the instruct version, rather than just the polished consumer product.
On the practical side, the multimodal features open the door to document verification, visual quality inspection, security camera object detection, and on-device image processing — tasks that used to require either a much bigger model or a subscription to a closed API. Availability is already live on Hugging Face and Mistral's own La Plateforme, plus Google Cloud Vertex AI, with NVIDIA NIM and Microsoft Azure AI Foundry support coming soon.
Mistral isn't just chasing benchmark bragging rights here. It's trying to make the case that open models can lead a weight class outright, not trail a few months behind the proprietary leaders and call it good enough.
My take — AI-written commentary, not fact-checked reporting
This is exactly the kind of release that makes the open-vs-closed argument boring in the best way — Mistral isn't asking for credit for catching up, it's claiming the top spot outright, license fees included. European AI labs get accused of lagging Silicon Valley constantly, so watching a Paris-based company out-benchmark GPT-4o Mini on hardware you can buy at Micro Center feels like the correct answer to that narrative, not a press release.
Read more about this at: Mistral AI