On March 16, Mistral released Small 4, a name that sounds unremarkable but hides a practical product idea—using the MoE architecture to merge three previously independent models into one, enabling a single deployment to handle all scenarios.
One Model, Three Uses
Mistral previously had three models, each with its own role:
- Magistral: Specializes in complex tasks requiring reasoning
- Pixtral: Handles multimodal image understanding
- Devstral: Dedicated to code generation and development
Using all three meant three deployments, three scheduling logics, and triple the operational costs. Small 4 merges them directly.
One model simultaneously supports text reasoning, image understanding, code generation, function calling, and JSON output, with a 256K context window.
The Core Logic of MoE: Big but Not Expensive
Small 4's parameter configuration is quite representative:
| Metric | Value |
|---|---|
| Total parameters | 119B |
| Number of experts | 128 |
| Experts activated per inference | 4 |
| Actual activated parameters | 6.5B |
| Context window | 256K |
119B looks intimidating, but each inference actually uses only 6.5B parameters—this is the core value of the MoE architecture: the model knows more, but only uses a small portion when computing. The inference cost is close to that of a 6.5B dense model, but the capability boundary is much wider.
Compared to Small 3, end-to-end latency has dropped by 40%, and the number of requests processed per second is three times higher.
This logic follows the same path as DeepSeek V3 and Qwen3: not relying on larger full parameters to impress, but using sparse activation to find a balance between efficiency and capability.
reasoning_effort: On-Demand Reasoning Capability
This is a practical aspect of the product design.
Previously, you had to choose between a fast-response model and a slow-reasoning model, often by deploying two different models. Small 4 adds a parameter that lets you control directly at the API request level: whether to use reasoning capability for this request, and how much.
Simple questions get a quick answer; complex questions get more thinking time—the same model, the same API, no switching. This design can save engineering teams a lot of trouble.
Open-Source License and Availability
Mistral continues to release under the Apache 2.0 license, with no commercial restrictions, and weights can be downloaded for self-deployment.
Currently available at:
- Mistral API and AI Studio
- Hugging Face (full weight download)
- NVIDIA build.nvidia.com (free prototype testing)
- NVIDIA NIM container (production deployment)
Apache 2.0 is the cleanest license for enterprises—no usage restrictions, can be modified, and can be distributed without issues.
In the Broader Open-Source Large Model Trend
In the past two years, the competitive landscape of open-source large models has changed rapidly. DeepSeek V3 used MoE plus low-cost training to erode the moat of closed-source models, Alibaba's Qwen3 opened up the landscape with the Apache license, and Meta's Llama 4 also released an MoE version in April.
Mistral's Small 4 follows a similar path: 119B total, 6.5B activated, Apache license, packaging reasoning, multimodal, and code all into one deployment unit. Two years ago, this configuration was only available in top-tier closed-source models; now it's free to download.
For teams looking to build AI capabilities in production, the choice of open-source models is shifting from "good enough" to serious consideration. Small 4's main value in this option pool is reducing deployment complexity—no need to maintain three models simultaneously; one is enough.
Whether it suits your specific scenario still requires running benchmarks. Mistral officially states that Chinese support is not as strong as English, which teams in China need to evaluate separately.
Sources: CocoLoop, 119B Parameters, 6.5B Activated: Mistral Small 4 Collapses Three Open Models Into One (BaristaLabs Blog); Mistral AI Releases Mistral Small 4 (MarkTechPost)