Mistral Small 4: One 119B MoE Model to Replace Four

Mistral AI released Mistral Small 4 today — a 119B-parameter Mixture-of-Experts model that consolidates what was previously four separate products into one. It replaces Mistral Small, Magistral, Pixtral, and Devstral. One model, one deployment, one API endpoint.

The architecture is 128 experts with 6B parameters active per token. That’s the core MoE tradeoff: 119B total parameters means strong reasoning capacity, but you’re only paying compute for 6B on any given forward pass. The context window is 256k tokens. Mistral claims 40% lower latency and 3x higher throughput compared to Mistral Small 3 — numbers that, if they hold in production, make this a serious option for anyone currently juggling multiple specialized endpoints.

The consolidation play is deliberate. Mistral has been fragmenting its model lineup for two years, releasing niche-specific variants to compete in specific benchmark categories. That strategy works for headlines but creates real operational friction: different API parameters, different fine-tuning paths, different latency profiles. Mistral Small 4 is a bet that a single capable generalist is more valuable to developers than a portfolio of specialists.

One genuinely useful feature is the reasoning_effort parameter. You can dial inference from none to high at request time, letting you trade latency for quality depending on the task. A classification call gets none. A complex code review gets high. This kind of controllable reasoning has been standard in OpenAI’s o-series but is new for Mistral. It’s a practical acknowledgment that not every call needs chain-of-thought.

The release is Apache 2.0 — full commercial use, no restrictions. It’s available on the Mistral API, Hugging Face, NVIDIA NIM day-0 containers, vLLM, and llama.cpp. The day-0 NVIDIA NIM support is notable: it means you can run Mistral Small 4 in optimized containers on H100s without waiting for third-party packaging.

Mistral’s open-source releases have consistently punched above their weight on benchmarks relative to model size, and the company’s Apache licensing has made it the default choice for enterprises that need to self-host. Whether a 119B MoE model is actually “small” enough for most self-hosted deployments is the real question — the memory footprint is significant even with MoE routing. But for teams running on multi-GPU infrastructure, the consolidation argument is compelling.

The timing is sharp. As OpenAI and Google push their models further behind API walls, Mistral is doubling down on openness as product differentiation.