On the open-source large model track, Mistral AI once again demonstrated what it means to "prioritize efficiency." Today, the company announced the official open-sourcing of its latest masterpiece in the Small series under the Apache 2.0 license — Mistral Small4, and simultaneously declared its membership in the newly established Nemotron alliance by NVIDIA, becoming one of the founding members.

This is not a simple routine upgrade. Mistral Small4 has been officially defined as the first "all-in-one" multi-functional model. It perfectly integrates the three flagship capabilities that the Mistral family has previously prided itself on:

Magistral: Excellent logical reasoning ability.

Pixtral: Strong native multimodal processing capability (supports image input).

Devstral: Professional intelligent agent coding ability.

This means developers no longer have to choose between "fast command models," "heavy reasoning engines," or "multimodal assistants" — one model can handle all scenarios.

In terms of technical specifications, Mistral Small4 adopts a precise 128-expert mixture-of-experts (MoE) architecture. Despite a total parameter count of 119 billion, through optimization, only 4 experts are activated per token (about 60 billion active parameters), significantly reducing computational costs while maintaining performance. In addition, the ultra-long context window of 256k allows it to easily handle long document analysis and complex conversations.

Notably, this model introduces a "configurable reasoning intensity" feature. Users can freely switch between low-latency fast responses and deep thinking reasoning outputs. Test data shows:

In the latency-optimized mode, the end-to-end completion time was reduced by 40%.

In the throughput-optimized mode, the number of requests per second (RPS) increased by 3 times compared to the previous generation.

From code completion to visual analysis, the open-sourcing of Mistral Small4 undoubtedly provides global developers with a more flexible and efficient "versatile foundation." When top-tier reasoning capabilities are combined with native multimodality, the innovation ceiling of the open-source community may be raised once again.