Ministral 3B and 8B

What it is

Ministral 3B and Ministral 8B are a pair of small language models released by Mistral AI on 16 October 2024, aimed at on-device and edge use rather than data-centre serving ^[1]. They were announced together and nicknamed "les Ministraux" ^[1]. The release date was the first anniversary of Mistral 7B, the earlier model from the same company ^[1].

The numbers in the names follow the usual convention of counting parameters in billions. Small models of this kind are meant to run close to the user — on a laptop, a phone, or a local machine — where memory, latency and the ability to work without sending data to a remote server matter more than raw scale. That is the setting the two models were built for ^[1].

Design

Both models accept context of up to 128K tokens, though at launch serving through vLLM supported 32K ^[1]. Long context in a model this small is the notable part: it lets a device-side assistant work through a long document or a lengthy conversation history without a larger model behind it.

Ministral 8B uses interleaved sliding-window attention ^[1]. In ordinary attention every token compares itself with every other token in the sequence, so cost grows quickly with length. Sliding-window attention instead lets each token look at a limited span of nearby tokens; interleaving mixes layers that use such windows with layers that do not, so information can still travel across the whole sequence. The point of the design here is speed and memory: it gives faster, more memory-efficient inference ^[1].

Benchmarks

Mistral said the two models consistently outperform their peers across multiple tasks ^[1]. The most striking claim was that Ministral 3B beats Mistral 7B on most benchmarks ^[1] — a smaller model from a later generation overtaking an earlier, larger one. Comparisons were made against Gemma 2 2B, Llama 3.2 3B, Llama 3.1 8B and Gemma 2 9B ^[1], which span both the small end of the field and the class above it. Mistral also noted that it re-evaluated all of the compared models with its own internal framework, so that the comparison ran under one consistent setup ^[1].

These are the company's own evaluations, published alongside the release; they indicate where Mistral placed the models rather than an independent ranking.

Pricing and licensing

On the API, Ministral 8B was priced at $0.10 per million tokens and Ministral 3B at $0.04 per million tokens, charged at the same rate for input and output ^[1]. The two also differed in licensing: Ministral 8B came under the Mistral Commercial and Research licences, with the Instruct weights available for research, while Ministral 3B was released under a commercial licence ^[1].

That split is worth noting because it decides who can do what with the weights — the smaller model was offered on commercial terms, while the 8B had a research route to its Instruct weights.

Retirement

Mistral deprecated the model names ministral-3b-2410 and ministral-8b-2410 on 2 December 2025 ^[2] and retired them on 31 December 2025, replacing them with the Ministral 3 family ^[2]. The pair therefore lived for a little over a year as served models, which is a normal lifespan for a small model in a fast-moving lineup.