Mistral Small 4 (26.03)

Mistral Small 4 is the Mistral open-weights model in the Small line, version 26.03, released on 16 March 2026. It has 119 billion parameters in total, 6.5 billion of them active for each token, and its weights carry the Apache 2.0 licence. The question we put to the reviews is the one in the topic: is running it at home cheaper than paying Mistral to run it?

We read the sources on 12 October 2026. The sample is three official pages, three editorial or evaluation pages (two published on 21 March and 11 June 2026, and one Artificial Analysis page read on 12 October), one guide from an API aggregator, one Hacker News thread of ten comments from 16 March, and one news report about a different, larger Mistral model. The editorial tier is thin and the owner tier is empty: no buyer platform reviews an open-weights model, and two pages we looked for could not be read.

Consensus

The sources agree on the facts of the model and do not agree on the question. They agree on the licence, the price and the size. They do not answer whether home or API is cheaper, because none of the nine adds up what a home setup costs, none adds up what the API costs for a stated workload, and none sets one against the other.

We read this as an unsettled question, not a settled answer. The free part is real: the weights are free to download under a licence file anyone can read. The part reviewers leave open is what it takes to run them. Confidence is strong for the listed facts, thin for the reviews of how it behaves, and absent for the cost comparison. The reading is of version 26.03 on 12 October 2026, and a model this young moves quickly.

Recurring strengths

The licence is open. Mistral's announcement, its model card and the Hugging Face card all state Apache 2.0, and so do all three editorial or evaluation pages (3 of 3). That is the maker's statement and the licence the repository shows, not a legal reading of what any reader may do with it.

The API price is low. The model card lists $0.15 per million input tokens and $0.60 per million output tokens, read on 12 October. The Artificial Analysis page lists the same two figures (1 of 3 editorial pages), and one of the ten Hacker News comments calls the output price low against rivals. These are the maker's list prices, not what any workload costs.

It keeps its answers short. Mistral says the model writes shorter outputs than a rival open model on the benchmarks it chose. That is a claim. Artificial Analysis reports 54 million output tokens used in its evaluation against a median of 100 million for comparable models, and AISO Tools repeats the efficiency claims, so 2 of 3 editorial pages describe it as concise. The same page puts its output speed at 165 tokens a second.

Recurring complaints

The hardware is server class. Mistral's announcement lists a minimum of four NVIDIA HGX H100, two HGX H200 or one DGX B200. AISO Tools repeats that list as the self-hosting requirement (1 of 3 editorial pages), and the CometAPI guide, published on 23 March and updated on 3 September, says the model is too heavy for consumer graphics cards without optimisation. A count of pages that list a requirement is not a count of owners who struggled, and no source we read reports an owner's experience of running it.

It trails the strongest models on hard tasks. AISO Tools says it does not match the top proprietary models on complex reasoning and research, and MindStudio says it falls short of a larger proprietary model on complex reasoning, so 2 of 3 editorial pages say so. One Hacker News commenter who ran it in an agentic workflow found it fine for basic tasks and not for complex ones. That is one developer's report.

Benchmarks may not predict use. One Hacker News commenter doubts that benchmark scores can be trusted when choosing between open models of about 120 billion parameters. It is one opinion in one thread.

Where reviewers split

What it takes to run at home. Mistral and the CometAPI guide put the floor at data-centre hardware. MindStudio says a quantized copy runs with 16 to 32 GB of memory or a graphics card with at least 16 GB. One Hacker News commenter puts a four-bit copy on a single H100 or a 128 GB machine. These differ by a wide margin, and we ran none of it. We read the maker's list as the only one that comes from the maker.

The context window. The model card lists 256 thousand tokens. MindStudio lists 128 thousand. We use the maker's figure, and we read the difference as a sign that one editorial page has details wrong.

Which route is cheaper. MindStudio says the economics favour self-hosting because there is no per-request fee. The price list and the hardware floor point the other way for a small user. Neither side gives a figure, so we do not call it.

A different model. The Register reported on 6 October that Mistral Large 4, a model of about one trillion parameters, was in preview by API with its weights still to come. That is a different model and is not counted anywhere above. We chose Small 4 because its weights are published and can be downloaded today.

Who it suits

You already run server-class graphics cards, or rent them, and want weights under a licence you can read. The licence is Apache 2.0 on three official pages.

You want to call a capable open model by API at a listed price and keep the option of moving the weights to your own machines later. The price is on the model card.

You want a model that activates 6.5 billion of its 119 billion parameters for each token, which Artificial Analysis lists as an output speed of 165 tokens a second.

Who should pass

You hoped to run it on one ordinary home computer. The maker's minimum, and a guide that calls it too heavy for consumer cards, point away from that, while one page disagrees.

You need the strongest reasoning on hard problems. Two of three editorial pages place it below the best proprietary models.

You were hoping a review would say which route costs less. None of the sources we read does, and we will not invent a break-even.

This is a reading of published reviews, not medical, financial or legal advice. For a decision about your health, your money or your rights, a qualified professional is the right next step, and not a review.

One verdict a week: the most useful Master Review we finished, the complaint that kept appearing, and who should skip it.

Sources

• Mistral, Small 4 announcement, official, published 16 March 2026.
• Mistral model card, Small 4, official, read 12 October 2026.
• Hugging Face model card, Mistral-Small-4-119B-2603, official, read 12 October 2026.
• Artificial Analysis, Mistral Small 4, editorial evaluation, read 12 October 2026.
• AISO Tools review, editorial, published 11 June 2026.
• MindStudio, what is Mistral Small 4, editorial, published 21 March 2026.
• CometAPI, how to run Mistral Small 4 locally, other (API aggregator guide), published 23 March 2026.
• The Register, on Mistral Large 4, news, a different model, published 6 October 2026.
• Hacker News thread, Mistral Small 4, forum, 16 March 2026.

On 12 October 2026 we read 3 official pages, 3 editorial or evaluation pages, 1 guide, 1 news report on a different model and 1 forum thread, published 16 March to 6 October 2026 or read on 12 October, by web search, one public API call and plain page fetch, and synthesised them with AI; nothing was run or tested.