Mistral Small 4: 119B Parameters, 6B Active, Europe's Efficient AI Play

Mistral released Mistral Small 4 on March 16, and the architecture is worth understanding before looking at the benchmarks. This is a Mixture of Experts model with 119 billion total parameters but only 6 billion active per token. That structure is what makes the performance numbers interesting.

The Architecture: Why MoE Matters

Mixture of Experts models don’t use all their parameters for every token. Instead, a routing layer decides which subset of “expert” subnetworks handles each token in a forward pass. For Mistral Small 4, 6 billion parameters are active per token out of a total pool of 119 billion.

The practical effect: inference cost and memory requirements are roughly proportional to the 6B active parameters, not the 119B total. You get reasoning quality that draws on a much larger total capacity than the inference cost implies. A 6B active MoE model doesn’t perform like a 6B dense model. The experts are drawn from a much larger pool, and the routing lets different experts specialize in different tasks.

Mistral has been refining this architecture for several generations. The routing in Small 4 has improved from previous models, with cleaner specialization between expert clusters and less redundancy.

Capability Profile: Instruction, Reasoning, Vision, Coding

Mistral Small 4 is built to handle four task types in a single model: instruction following, extended reasoning, visual inputs, and code generation. Previous Mistral Small variants had partial multimodal support; Small 4 fully integrates vision.

On instruction following, it performs well on standard benchmarks like MT-Bench and AlpacaEval. Scores land in the range of other mid-tier models, competitive with Gemini 3.1 Flash and ahead of many open-weight models in this parameter class.

On coding, HumanEval puts Small 4 around 82%, which is strong for a model with 6B active parameters. LiveCodeBench scores are more modest, around 38%, reflecting that harder multi-step problems favor models with larger active capacity. For typical code completion and explanation tasks, 82% on HumanEval means it handles the large majority of what developers actually ask of a coding assistant.

Vision is the newest addition to the lineup, and Small 4 handles standard image understanding benchmarks adequately, around 70% on MMMU, without reaching the quality of dedicated vision models or the largest multimodal flagships. It’s sufficient for applications that need image analysis alongside text without requiring a separate vision model in the stack.

The extended reasoning capability mirrors what other models have introduced: the model can think through problems step by step before producing a final answer. This is most useful for math and multi-step logic problems. On MATH-500, Small 4 scores around 89% in thinking mode.

The European Data Residency Angle

Mistral is a French company, and data residency in Europe is a genuine differentiator for European customers. GDPR and sector-specific regulations in finance, healthcare, and government create real requirements about where data is processed and stored. American AI providers offering EU data residency do so as an option, often at premium pricing, whereas Mistral runs on European infrastructure by default.

For a company subject to GDPR with strict data processing requirements, using a model whose infrastructure is natively European reduces compliance overhead. You’re not relying on a contractual arrangement that keeps data in EU data centers. You’re using a provider whose core operations are based there.

This matters most for European enterprise customers, regulated industries, and government organizations. For US-based developers building general-purpose applications, it’s less relevant. But for the organizations where it does matter, it’s not a minor consideration.

API Access and Open Weights

Mistral Small 4 is available through Mistral’s API at pricing competitive with Gemini Flash, around $0.10 per million input tokens at standard rates. The model is also available as open weights under the Apache 2.0 license, which permits self-hosting and commercial use. If self-hosting is the point, our guide to OpenCode local models with Ollama covers what it takes to drive a coding agent off a model on your own hardware.

Self-hosting is more demanding than the 6B active figure suggests. Sparse activation cuts compute, not memory: all 119B parameters have to stay resident for the router, which means roughly 119 GB of VRAM at 8-bit and around 240 GB in full precision. Mistral’s own guidance lists a minimum of 4x H100, 2x H200, or 1x DGX B200 for serving the model, and a single 80 GB H100 only fits it with aggressive quantization. This is enterprise-scale hardware, not a single-GPU deployment.

Where It Fits

Mistral Small 4’s niche is mid-tier capability with an efficient architecture and European compliance credentials. It’s not competing with Gemini 3.1 Pro or Claude Opus 4.6 on capability. It’s competing with Gemini Flash, Claude Sonnet, and open-weight models in the 7B-14B range.

For European developers who want a capable model they can self-host, audit, and keep within GDPR boundaries without negotiating data residency agreements with American cloud providers, Small 4 is currently the strongest option. For everyone else, it’s a competitive mid-tier choice worth benchmarking against the alternatives for your specific tasks.

Fitting the MoE architecture and a full capability profile (instruction, reasoning, vision, and coding) into a single model at this efficiency level is solid work. Mistral remains one of the more interesting players for developers who care about deployment flexibility as much as raw benchmark numbers.