Mistral Small 3.1 adds multimodal input and a 128k context window
Mistral AI
- Models
- Open source
- #mistral small 3.1
- #apache 2.0
- #multimodal
Mistral announced Mistral Small 3.1, an update to Mistral Small 3 with improved text performance, multimodal understanding and a context window of up to 128k tokens. Mistral says it outperforms Gemma 3 and GPT-4o Mini and reaches 150 tokens per second, and it ships under Apache 2.0 with base and instruct checkpoints.
What Mistral announced
Mistral announced Mistral Small 3.1 as an update to Mistral Small 3, saying it improves text performance, adds multimodal understanding and expands the context window to up to 128k tokens. The company positions it as outperforming comparables such as Gemma 3 and GPT-4o Mini while delivering inference at 150 tokens per second, describing these as its own or previously reported figures. It is released under Apache 2.0, with both base and instruct checkpoints, and Mistral lists uses including image understanding, function calling and fine-tuning for specialized domains. The model is offered for download on Hugging Face, via Mistral's La Plateforme API and on Google Cloud Vertex AI, with subsequent availability on NVIDIA NIM and Microsoft Azure AI Foundry. Mistral says it can run on a single RTX 4090 or a Mac with 32GB RAM.
A scenario for a platform engineer
A platform engineer at a company processing scanned supplier invoices could consider trialling Mistral Small 3.1's image understanding and long context together, feeding a multi-page document plus extraction instructions in one request and checking whether returned fields match the source. Because the model is offered both as downloadable checkpoints and through hosted APIs, the engineer could compare a locally hosted instruct checkpoint against an API route on the same sample set. This is a hypothetical trial design, not a demonstrated outcome; the announcement provides no evidence about invoice accuracy.
How to run a bounded trial
A team could assemble a fixed set of real tasks, record the current baseline (accuracy, latency, cost per request), then run the same set through Mistral Small 3.1 and compare results. Include failure checks that target the new capabilities specifically: multi-page cases that exceed the old context limit, image-plus-text prompts, and function-calling steps that need to complete correctly. Measure whether gains come from the update or from prompt changes by keeping prompts identical. Keep a human reviewer in the loop for consequential outputs, for example checking extracted figures before payment. Define in advance what result would count as no improvement.
Trade-offs and open questions
The longer context and image support widen the range of tasks a small model can address, but longer prompts can raise compute needs, and Mistral's hardware guidance refers to running the model generally rather than to peak memory at 128k tokens. The speed and comparison figures come from Mistral, so a team should verify them on its own workload rather than relying on the announcement. Similarly, fine-tuning for regulated fields such as legal or medical advice presumes data, evaluation and compliance work that the announcement does not cover. A team should avoid switching production systems based on this excerpt alone, and should first confirm whether the hosted options fit its data-residency and licensing requirements.
Context and limitations
- Based on a short excerpt of the announcement; the full article and benchmark tables were not analyzed.
- Performance, speed and comparison claims come from Mistral and were not independently verified.
- The excerpt gives no pricing, latency baselines for specific tasks, or dates for NVIDIA NIM and Azure AI Foundry availability.
- Based on bounded source text, not independent testing. Practical suggestions are editorial proposals.