# Mistral Small 3 made latency a claim to measure in the target stack Mistral introduced Mistral Small 3 on January 30, 2025 as a 24B Apache 2.0 model optimized for low latency. A rate claim needs a reproducible serving profile. Canonical: https://rangoon.ai/insights/archive-mistral-small-3-latency-2025/ Author: Rangoon Editorial (https://rangoon.ai/insights/editorial/) Published: 2026-10-23 Updated: 2026-10-23 Event date: 2025-01-30 Tags: [Mistral](https://rangoon.ai/insights/tags/mistral/), [Mistral Small 3](https://rangoon.ai/insights/tags/mistral-small-3/), [latency](https://rangoon.ai/insights/tags/latency/), [open weights](https://rangoon.ai/insights/tags/open-weights/) Source note: Historical reporting: Mistral introduced Mistral Small 3 on January 30, 2025. Mistral’s model documentation records the version and Apache 2.0 status; analysis below is editorial. Image: Mistral Small 3: latency budget. Original Rangoon explanatory diagram; conceptual relationships, not an event photograph, vendor interface, or measured result. ## Key takeaways - Mistral introduced Small 3 as a latency-optimized 24B-parameter model under Apache 2.0. - A published tokens-per-second result depends on hardware, runtime, batch policy, prompt length, and output length. - Serving evaluation should separate time-to-first-token, decode rate, tail latency, and queue delay. ## What Mistral introduced [source 1](https://mistral.ai/news/mistral-small-3/) [source 2](https://docs.mistral.ai/models/mistral-small-3-0-25-01) On January 30, 2025, Mistral introduced Mistral Small 3 as a latency-optimized 24B-parameter model released under the Apache 2.0 license. Its announcement described the model as aimed at language and instruction-following tasks with low latency, and reported benchmark and throughput comparisons under its stated conditions. The post also positioned the size as suitable for local deployment. Mistral’s model documentation identifies Mistral Small 3.0 v25.01 as a 24B open model, notes the January 30, 2025 blog date, and lists Apache 2.0. The documentation gives a versioned product record. It does not establish a particular organization’s serving speed, total operating cost, or the suitability of a later variant for a different task. ## Editorial analysis: one throughput number hides a serving path Tokens per second is useful only when paired with the conditions that produced it. An interactive user experiences time to first token, token-by-token decode time, queueing, retrieval delay, tool calls, network transport, and rendering. A batch system may value aggregate throughput instead. Reporting one maximum decode rate as the experience of both workloads obscures the design choice. Build a serving profile that names model artifact, runtime, precision, accelerator, context length, input length distribution, requested output length, concurrency, batching policy, cache state, and network boundary. Capture median and tail latency separately. This lets a team see whether a slowdown came from model decoding, a long prompt, an overloaded queue, or a dependency outside the inference engine. ## Editorial analysis: open weights still need artifact control An Apache 2.0 model release can simplify one licensing dimension, but deployment still needs a concrete artifact decision. Record the source location, checksum, revision, conversion or quantization steps, runtime image, and any added prompt template. These details determine what the service actually runs and make it possible to investigate a result after the upstream repository changes. Use an approved evaluation corpus that reflects the intended task, then measure correctness and operational behavior together. A fast completion with broken JSON, unsupported citations, or a poor routing decision does not satisfy an interactive service. Conversely, a slower but reliable configuration may be appropriate for asynchronous work with a different service-level objective. ## Editorial analysis: make capacity a planned decision Set capacity targets before rollout: concurrent requests, target first-token latency, target completion latency, maximum queue time, memory headroom, and failure behavior when the target is exceeded. Exercise the overloaded path with the same model artifact and prompt distribution used for the baseline. A system that is fast for one isolated request can still fail users when requests arrive together. Mistral Small 3’s release is a historical example of latency being a product claim. The operational lesson is to preserve a measured serving profile beside that claim, so teams can compare configurations and choose a workload fit without treating a vendor number as a universal result. The same profile makes a capacity regression visible before it becomes a user-facing queue. ## Sources 1. [Mistral: Mistral Small 3](https://mistral.ai/news/mistral-small-3/) 2. [Mistral Docs: Mistral Small 3.0](https://docs.mistral.ai/models/mistral-small-3-0-25-01) ## Related reading - [Mistral Small 3.1 made multimodal acceptance a modality-by-modality test](https://rangoon.ai/insights/archive-mistral-small-3-1-2025/) - [Tencent released Hunyuan-A13B under a territory-limited community license, separating open weights from universal permission](https://rangoon.ai/insights/archive-tencent-hunyuan-a13b-open-hybrid-reasoning-june-2025/) - [GLM-4.7 made preserved thinking a named control for multi-turn coding agents](https://rangoon.ai/insights/archive-zai-glm-4-7-december-2025/) ## More information - [Documentation index](https://rangoon.ai/llms.txt) - [AI access and policies](https://rangoon.ai/ai/) - [Source repository](https://github.com/hypler-dev/rangoon)