Published on March 20, 2026
Mistral Small 3.1 Makes Way for Mistral Small 4
Mistral has released Small 4, a new compact model that's faster, more accurate, and boasts improved performance across multiple languages, including Russian.
Products 4 – 5 minutes min readCompact language models aren't «stripped-down AI» but rather fast, cost-effective, and sufficiently intelligent solutions for specific tasks that don't require bringing in the «heavy artillery.» Mistral has long been working in this direction and has now introduced the next generation of its compact lineup – Mistral Small 4.
What is Mistral Small?
In short, it's a series of relatively small language models from the French company Mistral AI, positioned as a practical alternative to large models where speed and cost are key. They don't claim to be the most powerful, but they aim for maximum utility with minimal resources.
Mistral Small 4 is the latest release in this series. The model has 22 billion parameters and, according to its developers, delivers performance comparable to larger competitors at this size.
What Has Changed Compared to the Previous Version?
The main focus for Mistral is handling long texts. Mistral Small 4 supports a context window of 128,000 tokens. Simply put, the model can «keep in mind» a very long conversation or a large document – much like you could give an AI an entire book to read and then ask questions about its content.
Multilingual capabilities have also been significantly improved. The developers specifically highlight French, German, Spanish, Italian, Portuguese, Arabic, Hindi, Russian, Chinese, and Japanese. This means the model now better understands and generates text in these languages – it doesn't just «know them» but works with them at a higher quality.
Another significant update is image support. Mistral Small 4 has become multimodal: it can not only read text but also analyze images. This opens up a whole new range of applications – from describing images to answering questions about visual content.
How Does It Perform in Practice?
On benchmarks (standard tests for comparing models), Mistral Small 4 looks strong in its class. Its developers compare it to GPT-4o mini and Gemma 3, claiming it surpasses both models on several metrics.
Tasks related to coding and reasoning are also worth mentioning. Here, the model also shows significant progress compared to the previous generation, Mistral Small 3.1.
However, it's important to understand that benchmarks are synthetic tests. The model's real-world utility depends on the specific scenario. But overall, the numbers confirm that Mistral Small 4 is a meaningful step forward, not just a change in the version number.
Who Will Find This Useful?
Mistral Small 4 is primarily aimed at developers and companies looking to integrate language AI into their products. Its compact size allows it to run with lower computational costs, which is crucial when dealing with high volumes of requests or running on proprietary hardware.
The model is available via the Mistral API under the name mistral-small-2503, and it is also open-access – its weights are published on the Hugging Face platform under the MistralAI Research License. This means researchers and enthusiasts can download the model and run it locally.
For those who just want to try it out – without any installation or setup – it's possible to interact with the model directly through Le Chat, Mistral's proprietary chat interface.
Why Is This Interesting in the Context of the Industry?
Over the past few years, an interesting trend has emerged in the AI industry: companies are racing to build ever-larger models, but at the same time, interest in compact solutions is growing. The reason is simple: large models are expensive to maintain, and for most real-world tasks, their full power is simply not needed.
Mistral occupies a clear niche in this space: they are betting on efficiency. Mistral Small 4 is a great example of how to achieve a lot with relatively modest resources. 22 billion parameters isn't a giant model by modern standards, but with support for a long context, images, and multiple languages, it covers a fairly wide range of tasks.
The open publication of its weights is also important: it allows the community to adapt the model for their own needs – fine-tuning it on specific data, optimizing it for particular hardware, and using it in closed environments without transmitting data to third parties.
So, Mistral Small 4 is not just another release. It's further proof that in the world of AI, «smaller» doesn't necessarily mean «worse.»
You May Also Like
Events are only part of the bigger picture. These materials help you see more broadly: the context, the consequences, and the ideas behind the news.
The new version of Mistral's multimodal model can work with images, video, and text simultaneously, doing so quickly and accurately.
The Bangalore-based developer has released a multimodal model that understands speech, text, and images, supports India's major languages, and is capable of operating offline.
AI: Events
Products
Mixedbread has introduced Wholembed v3 – a multimodal search model that works with text and images in dozens of languages and claims to deliver best-in-class results.
How This Text Was Created
This material is not a direct retelling of the original publication. First, the news item itself was selected as an event important for understanding AI development. Then a processing framework was set: what needs clarification, what context to add, and where to place emphasis. This allowed us to turn a single announcement or update into a coherent and meaningful analysis.
We openly show which models were used at different stages of processing. Each performed its own role — analyzing the source, rewriting, fact-checking, and visual interpretation. This approach maintains transparency and clearly demonstrates how technologies participated in creating the material.
The neural network studies the original material and generates a coherent text
Correction of errors, inaccuracies, and ambiguous phrasing
Generating a textual prompt for the visual model
Generating an image based on the prepared prompt
Want to dive deeper into the world
of neuro-creativity?
Be the first to learn about new books, articles, and AI experiments
on our Telegram channel!
