Mistral Small 4 Launches: A Powerful, Efficient Three-in-One Open Source Model

In the open-source large language model space, Mistral AI has once again proven its commitment to "efficiency first." Today, the company officially open-sourced its latest Small series flagship, Mistral Small4, under the Apache 2.0 license, while announcing its role as a founding member of NVIDIA’s newly launched Nemotron alliance.
This goes beyond a standard update. Mistral Small4 is officially positioned as the first "all-in-one" multimodal model, seamlessly combining three core strengths previously offered by separate Mistral models:
Magistral: Advanced logical reasoning.
Pixtral: Native multimodal processing (including image understanding).
Devstral: Specialized coding and agent capabilities.
Developers no longer need to choose between speed, deep reasoning, or multimodal features—this single model handles all use cases.
Technically, Mistral Small4 utilizes a precise 128-expert mixture-of-experts (MoE) architecture. With 119 billion total parameters, it activates only 4 experts per token (roughly 60 billion active parameters), drastically cutting compute costs without sacrificing performance. It also supports a 256k context window, enabling seamless handling of long documents and complex multi-turn conversations.
Notably, the model features "configurable reasoning intensity," allowing users to toggle between low-latency responses and deep-thinking outputs. Benchmarks show:
In latency-optimized mode, end-to-end completion time dropped by 40%.
In throughput-optimized mode, requests per second (RPS) increased by 3x compared to the previous generation.
From code generation to visual analysis, the open-source release of Mistral Small4 offers global developers a more flexible and efficient "universal foundation." By combining top-tier reasoning with native multimodality, it may once again raise the bar for open-source innovation.
Related article
Librarians Host Viral 'Avoiding AI' Workshops For Those Fed Up With Big Tech
“Everyone is glued to their screens in my library program,” joked Charlie Bailey, a South Philadelphia librarian. He invited his audience to take out their phones so he could guide them through disabling Apple Intelligence and Gemini.Bailey stands in
Anthropic’s Safety Warnings Backfire as Government Halts Its Most Powerful AI
On Friday, the U.S. government mandated that Anthropic immediately disable access to two of its most advanced AI systems, Claude Fable 5 and Claude Mythos 5, citing national security risks. Anthropic confirmed compliance via X but stated it believes
Adobe Unveils AI Assistant for Photoshop
Adobe announced on Tuesday that its AI assistant for Photoshop is becoming available to users in beta on the web and in the mobile apps. The company is also adding new AI-powered image editing capabilities to Firefly, its tool for media generation an
Related Special Topic Recommendations
Comments (0)
0/500

In the open-source large language model space,
This goes beyond a standard update. Mistral Small4 is officially positioned as the first "all-in-one" multimodal model, seamlessly combining three core strengths previously offered by separate Mistral models:
Magistral: Advanced logical reasoning.
Pixtral: Native multimodal processing (including image understanding).
Devstral: Specialized coding and agent capabilities.
Developers no longer need to choose between speed, deep reasoning, or multimodal features—this single model handles all use cases.
Technically, Mistral Small4 utilizes a precise 128-expert mixture-of-experts (MoE) architecture. With 119 billion total parameters, it activates only 4 experts per token (roughly 60 billion active parameters), drastically cutting compute costs without sacrificing performance. It also supports a 256k context window, enabling seamless handling of long documents and complex multi-turn conversations.
Notably, the model features "configurable reasoning intensity," allowing users to toggle between low-latency responses and deep-thinking outputs. Benchmarks show:
In latency-optimized mode, end-to-end completion time dropped by 40%.
In throughput-optimized mode, requests per second (RPS) increased by 3x compared to the previous generation.
From code generation to visual analysis, the open-source release of Mistral Small4 offers global developers a more flexible and efficient "universal foundation." By combining top-tier reasoning with native multimodality, it may once again raise the bar for open-source innovation.
Librarians Host Viral 'Avoiding AI' Workshops For Those Fed Up With Big Tech
“Everyone is glued to their screens in my library program,” joked Charlie Bailey, a South Philadelphia librarian. He invited his audience to take out their phones so he could guide them through disabling Apple Intelligence and Gemini.Bailey stands in
Anthropic’s Safety Warnings Backfire as Government Halts Its Most Powerful AI
On Friday, the U.S. government mandated that Anthropic immediately disable access to two of its most advanced AI systems, Claude Fable 5 and Claude Mythos 5, citing national security risks. Anthropic confirmed compliance via X but stated it believes
Adobe Unveils AI Assistant for Photoshop
Adobe announced on Tuesday that its AI assistant for Photoshop is becoming available to users in beta on the web and in the mobile apps. The company is also adding new AI-powered image editing capabilities to Firefly, its tool for media generation an





Home






