119Bなのに実質6.5B！Mistral Small 4が示すOSS LLM新基準Mistral Small 4 achieves effective inference at roughly 6.5B active parameters…

Mistral Small 4 is an open-source large language model that pairs a 119-billion-parameter footprint with a Mixture-of-Experts (MoE) design that activates only about 6.5 billion parameters during inference. That contrast is the headline: it aims to deliver the capacity of a very large model while keeping the per-token compute closer to that of a mid-sized one. For anyone trying to run capable models locally rather than through a hosted API, that trade-off is the central practical question, which is why the release is being framed as a possible new reference point for efficient open-weight models.

The mechanism behind the numbers is worth unpacking. A conventional dense model uses every one of its parameters to process every token, so a 119B dense network would demand roughly 119B parameters' worth of computation on each step. An MoE model instead splits parts of the network—typically the feed-forward layers—into many specialized subnetworks called experts, and a small routing component decides which experts handle a given token. Only the selected experts fire, so the model computes with a fraction of its total weights at any moment. In Mistral Small 4's case, the 119B figure describes the total stored capacity, while the roughly 6.5B active-parameter figure describes how much of that capacity is engaged per token.

This distinction between total and active parameters has direct consequences for how the model behaves in practice. Compute cost and latency scale mainly with the active parameters, so inference can feel far lighter than the 119B label suggests. Memory, however, tends to track the total parameter count, because all experts generally need to be available to be routed to, even if most stay idle on any given token. In other words, a model like this is likely to be relatively fast per token but still hungry for RAM or VRAM to hold the full set of weights. That is a familiar tension in MoE deployment, and it shapes which hardware setups can realistically host the model, especially in a llm">local-LLM context where memory is often the binding constraint.

Mistral's use of MoE is not new territory for the company. Its earlier Mixtral models, released as sparse Mixture-of-Experts networks, helped popularize the approach in the open-weight community by showing that routed experts could rival denser models at lower active cost. The broader industry has moved in the same direction: several prominent open and semi-open releases, including work from DeepSeek and Alibaba's Qwen line, have leaned on MoE to push total capacity upward without a proportional rise in inference compute. Seen against that backdrop, Mistral Small 4 appears to continue an established trajectory rather than introduce an entirely novel paradigm, and the claimed efficiency should be read as part of that trend.

Prospective users will also care about the surrounding tooling, since an open-weight model is only as useful as the runtimes that can serve it. Local inference stacks such as llama.cpp, Ollama, and LM Studio, along with higher-throughput serving frameworks like vLLM, are the usual paths for running models of this kind, and MoE support in these tools has matured considerably. Quantization—reducing the numerical precision of weights—remains a common way to shrink memory demands, and it is likely to be relevant here given the large total parameter count. How well a specific quantization level preserves quality tends to vary by model, so real-world performance is best confirmed through direct testing rather than assumed from the parameter figures alone.

A few caveats are worth keeping in mind. The source material presented here centers on the total-versus-active parameter story and positions the model as a new standard for efficient open-source LLMs, but it does not, in this excerpt, provide detailed benchmark comparisons, licensing specifics, or context-length figures. Readers evaluating the model for production or research should verify those details from primary documentation. The efficiency argument is compelling on paper, yet MoE models can be sensitive to routing behavior and workload mix, and their advantages are strongest when active-parameter savings are not offset by memory or serving overhead. Treated carefully, Mistral Small 4 illustrates how the open-model ecosystem is increasingly using sparsity to reconcile scale with the constraints of local deployment.