Mistral Small 4 Is One Model That Does the Job of Three — and It's Fully Open Source
Mistral's new 119B Mixture-of-Experts model unifies reasoning, multimodal understanding, and agentic coding under a single Apache 2.0 license. With only 6B active parameters per token and a 40% latency reduction in optimized configs, it's the most capable open-weight model Mistral has shipped.
Mistral dropped Small 4 on March 16 and it is a genuine architecture shift for the company. The model packs 119 billion total parameters into a Mixture-of-Experts framework with 128 specialized experts, but activates only four of them — 6 billion active parameters — per forward pass. The practical effect: a model with the capability surface of a 119B system running at the inference cost of a 6B system.
That trade-off is why MoE is winning.
The bigger story here is consolidation. Until Small 4, Mistral ran three separate specialized models: Magistral for reasoning, Pixtral for multimodal tasks, and Devstral for agentic code generation. All three had distinct deployment requirements. Small 4 absorbs all of them. One checkpoint, one hardware footprint, one API call. The reasoning_effort parameter lets developers dial the model between fast-and-shallow and slow-and-thorough without changing models.
License is Apache 2.0. Full weights, no usage restrictions, fine-tunable on proprietary data. For enterprises that cannot route sensitive data through closed APIs, this is a meaningful unlock. The requirement is 4 Nvidia HGX H100 GPUs minimum — expensive but within reach for any serious ML team.
Performance specifics: in latency-optimized configurations, Small 4 reduces end-to-end completion time by 40% compared to running separate Magistral/Pixtral/Devstral deployments. On the Artificial Analysis coding leaderboard, it sits above GPT-5.4 Standard in coding tasks while remaining below Claude Opus 4.6 on pure reasoning benchmarks. It processes text and images natively with no modality switching.
Mistral also shipped Forge alongside Small 4 — an enterprise platform for training custom versions of its models on proprietary datasets. Think fine-tuning infrastructure as a managed service. That is a different product from the model itself, but it clarifies Mistral’s commercial ambition: not just releasing open weights and hoping developers adopt them, but building a pipeline from open model to enterprise fine-tune to production deployment.
The competitive pressure behind this release is obvious. Meta’s Llama 4 is expected within weeks. Google’s Gemma 4 is already available. In a market where open-weight models are converging on capability, consolidation — fewer deployments covering more use cases — becomes a competitive moat. Developers who standardize on Small 4 for all three of Mistral’s original verticals are sticky in a way that one-use-case model users are not.
Mistral is a French company that has been consistently and credibly open-source in a way its larger competitors have not. Small 4 is the clearest expression of that thesis: maximum capability, minimum deployment friction, zero licensing barriers.