Mistral AI releases Ministral 3B and 8B for edge devices

What happened

On 16 October 2024, Mistral AI released two language models, Ministral 3B and Ministral 8B, aimed at on-device and edge computing ^[1]. The release date also marked the first anniversary of Mistral 7B ^[1].

The names state the essentials: a 3-billion-parameter model and an 8-billion-parameter model. Parameters are the trained numbers inside a network, and their count is a rough proxy for how much memory the model needs and how much computation each answer costs. By the standards of data-centre models these are small, and small is the point — a model of this size can in principle run on hardware a user already owns.

Why "edge" changes the constraints

On-device and edge computing describe where a model runs rather than what it can do: on the user's own machine — a laptop, a phone, a browser, or a small server sitting close to the data — instead of in a remote data centre reached over a network ^[1]. That setting changes the engineering constraints. Memory, battery, storage and cooling are all limited, so parameter count matters directly. It also changes what the deployment buys: because requests do not have to travel to a server, responses can start sooner, the system can keep working when connectivity is poor, and text can stay on the local device instead of being sent outwards.

Key details

Both models support context windows of up to 128K tokens ^[1]. The context window is the stretch of text a model can hold in view while answering — the conversation so far, or the document being read — and everything beyond it has to be summarised, retrieved separately, or dropped. A window of this size is what makes long inputs, such as lengthy documents or extended exchanges, workable without constant trimming.

Mistral also priced the two models through its API at $0.04 per million tokens for Ministral 3B and $0.10 per million tokens for Ministral 8B ^[1]. Per-million-token pricing is the usual way API access is sold: customers pay for text passing in and out rather than for a licence, which makes the cost of a workload proportional to its volume. Publishing the two figures side by side lets a developer see what it costs to route work to the smaller model instead of the larger one.

The company made the weights of the 8B Instruct version available for research use ^[1]. Weights are the trained parameters themselves; handing them out lets researchers run and study the model on their own machines rather than only through an API. The qualifier matters — research use is narrower than an open-source licence, and it is the 8B model, not the 3B, that is covered.

How the comparison was made

Mistral re-evaluated all of the models it compared against using its own internal evaluation framework, in order to make the comparison fair ^[1]. This is a real methodological point. Benchmark scores published by different organisations are produced with different prompts, different sampling settings and different scoring code, so headline numbers that look directly comparable often are not. Re-running every model under one framework removes some of that noise and puts the models on the same footing.

The trade-off is the familiar one: when the evaluation harness and the reported results both come from the company releasing the models, the numbers are best read as the vendor's own measurements rather than as an independent verdict. Whether the small models hold up on a given task still depends on testing them on that task.

What it signals

The release pairs a small size with a long context window and low per-token prices, and it arrived exactly one year after Mistral 7B ^[1]. The pattern it illustrates is common among model developers: rather than shipping a single flagship, offer a family at several sizes so that users can choose where to sit on the trade-off between cost and capability. For edge deployment specifically, the useful question is not whether a small model matches a large one, but whether it is good enough for the job while fitting the hardware it has to run on.