Mistral Small 4 Drops: 119B Parameter Open-Source MoE Model Under Apache 2.0

Mistral AI has released Mistral Small 4, a hybrid multimodal mixture-of-experts (MoE) model with 119 billion total parameters and just 6 billion active parameters per token. Released under the Apache 2.0 license, it unifies instruct, reasoning, and coding capabilities in a single open-source package that can be freely fine-tuned and self-hosted.

The model supports both text and image inputs and has been benchmarked as competitive with—or outperforming—closed-source models three to five times its effective compute size on reasoning and instruction-following tasks. The MoE architecture keeps inference costs manageable despite the large total parameter count, since only a fraction of the model activates for each token.

Mistral Small 4 is available on Hugging Face and through Mistral’s API, with the company also launching Mistral Forge for enterprise deployment and fine-tuning workflows.

Source

Mistral AI Blog | OpenRouter | Silicon Republic

Why This Matters

Mistral continues to be the most credible open-source challenger to the big labs. The “Small 4” naming is almost tongue-in-cheek at 119B parameters, but the MoE architecture means you’re only running 6B active params per token—making it genuinely deployable on reasonable hardware. Apache 2.0 licensing means no usage restrictions, no revenue caps, no “open but not really” shenanigans.

What’s notable is the convergence: multimodal, reasoning, and coding in a single model under a permissive license. A year ago, you needed separate specialized models for each capability. Now you can self-host a single model that handles all three, fine-tune it on your data, and deploy it without paying per-token API fees. For companies with privacy requirements or cost sensitivity, this is the kind of release that makes the build-vs-buy calculation shift meaningfully toward build.