Mistral has just released Mistral Small 3.1, a model that sets a new benchmark for performance in its class. Building on the success of Mistral Small 3, this latest version offers improved text capabilities, enhanced multimodal understanding, and an expanded context window of up to 128k tokens. Despite its relatively compact size (24B parameters), it outperforms competitors like Gemma 3, GPT-4o Mini, and Cohere Aya-Vision in multiple benchmarks. It also boasts impressive inference speeds of 150 tokens per second.

Key Features and Capabilities

1. Compact & Efficient: Small 3.1 can run seamlessly on an RTX 4090 or a Mac with 32GB RAM, making it ideal for on-device use.
2. Fast & Responsive: Delivers quick, accurate replies for virtual assistants and real-time applications.
3. Low-Latency Execution: Enables smooth function calling for automated and agent-based workflows.
4. Domain Adaptability: Fine-tunes for specialized fields like legal, medical, and technical support.
5. Advanced Reasoning: Supports in-depth problem-solving, with open checkpoints for further customization.
6. Broad Applications: Excels in document verification, diagnostics, image analysis, security monitoring, and customer support.

Performance Highlights

Mistral Small 3.1 has been rigorously tested across a variety of benchmarks, and the results speak for themselves. Here’s how it stacks up against other models in its class:

1. Text Benchmarks: In tasks like SimpleQA and GPQA Diamond, Mistral Small 3.1 consistently outperforms competitors, including Gemma 3-it and GPT-4o Mini with 10.43% and 44.42% score grade.
2. Multimodal Benchmarks: The model excels in multimodal tasks, achieving high scores in benchmarks like MMMU-Pro (49.25%), MathVista (62.8), and ChartQA (86.24).
3. Long Context Tasks: With its expanded context window, Mistral Small 3.1 performs exceptionally well in long-context benchmarks like LongBench v2 (37.18) and RULER 128k (93.96).

Availability and Deployment

Mistral Small 3.1 can be downloaded from:

huggingface website: https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Base-2503

https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503

Mistral AI’s developer playground: https://mistral.ai/news/la-plateforme

Google Cloud Vertex AI: https://cloud.google.com/vertex-ai/generative-ai/docs/partner-models/mistral.

With this release, Mistral continues to provide an open-source alternative that balances performance, accessibility, and efficiency. For those looking to build AI-driven applications without the overhead of larger proprietary models, Mistral Small 3.1 presents a compelling option.