S1-mini: the declared licence has the file, and the link works
7 min read
superwhisper/s1-mini appeared on Hugging Face on 12 August 2026: createdAt 2026-08-12T14:33:55Z, last modified 2026-08-21T17:16:30Z. We queried it on 22 August: 1,913 downloads, 198 likes, 12 files, a single .safetensors shard. Tags: transformers, qwen3, text-generation, asr, text-normalization, speech-to-text, dictation, license:other. Who is Superwhisper? The organisation behind the dictation app of the same name for Mac; 2 users, 2 models, 1 Space, 82 followers, not verified (isVerified: false) — smaller, on every metric, than the organisations behind the three cases we checked over the previous three days.
A corrector, not a chatbot, and it says so plainly
S1-mini doesn’t generate free text: it takes a raw transcript from an ASR engine — Whisper, Parakeet or another — and rewrites it clean, fillers removed, self-corrections resolved, punctuation and capitalisation applied according to a register chosen by whoever integrates it. The card states this without ambiguity: “S1-mini is not a chat model and will not follow general instructions; it does one job.” config.json confirms Qwen3ForCausalLM, a fine-tune of Qwen3-0.6B: 28 layers, 16/8-head GQA, hidden_size 1024, context up to 40,960 tokens inherited from the base — though the card recommends staying under 1,000 tokens of input, the task it was trained for.
On hardware the card is explicit, an exception among cards that say nothing at all: “the quantized build is a 462 MiB file that runs comfortably on a laptop CPU.” The parameter count is corrected by hand: the Hub sidebar shows 0.8B, but the text explains that the 155.6M-parameter embedding is counted twice, for 751.6 million tensor elements against 596.0 million unique parameters — the API confirms the first figure: 751,632,384. Evaluated on 7,519 cases in English, it declares 94.8% token accuracy, measured on the quantised GGUF Q4_K_M build and not on the BF16 weights in this repository: only the latter “should do at least as well”, with no figure of its own.
The declared licence has the file, where the link says it does
Here comes the line-by-line check — the one that, over the previous three days, found a gap three times running. The public tag is license:other; the YAML front matter adds license_name: s1-mini-license and license_link: LICENSE, a relative path, not an absolute URL as in Ornith’s case. We resolved it: https://huggingface.co/superwhisper/s1-mini/raw/main/LICENSE returns HTTP 200. The file exists and is exactly what the link promises. The other variants the platform uses for this document — LICENSE.md, LICENCE, NOTICE, USE_POLICY.md, ACCEPTABLE_USE.md, TERMS.md — 404 on all six: there isn’t a missing second restricting document, there simply isn’t one. gated is false, verified without authenticating: API 200, web page too.
The file is the Apache License 2.0 in full, followed by a clause that amends it: “any use, distribution, or integration of this model […] must continue to identify it by its original name, ‘S1-mini’ by ‘Superwhisper’, using that exact capitalization, regardless of any other name under which the model or a product incorporating it is marketed or distributed.” A single added condition, declared in the licence’s own name (s1-mini-license), not buried in a footnote.
This isn’t an isolated case in the positive sense, after three days of the opposite: on 19 August, Ant Group’s Ling-3.0-tiny, licence declared and file absent; on 20 August, Tencent’s UI-Mate-27B, the same pattern; on 21 August, Ornith-1.5-397B, with a link to a specific file that still returned 404. We found a fourth case today too, by way of contrast: sensenova/SenseNova-U1.5-8B-MoT, published on 19 August by the organisation tied to SenseTime (not verified, 37 users, 35 models, 504 followers), declares license:apache-2.0 in its tags but has none of the five licence files — the same 404 across the board. We won’t recount it in detail: the difference, today, is that the alternative exists.
Even here, one promise that doesn’t hold
The card carries a clickable badge: “Releases are tagged, so you can pin one,” with the instruction revision="v1" and a link to /tree/v1. We opened it: HTTP 404, on the page and on raw. The references listed via the API (/refs) confirm it: no tags, a single branch, main. The sibling repository superwhisper/s1-mini-GGUF, created two hours later, carries its own LICENSE, verified 200; the other links the card promises — the website, Discord, Qwen/Qwen3-0.6B, the arXiv paper — all return 200. Three Spaces already built on the model within ten days, including the organisation’s own (Gradio, tagged mcp-server, 200): the same model, no extra capability beyond the weights — unlike other cases we’ve checked, there is no hosted version here with functions the weights lack. A licence done well isn’t a perfect licence: there’s still a badge promising a tag that doesn’t exist. But here the problem is cosmetic, not the missing file.
The perimeter, even when the model is small
Downloadable weights without the compute that makes them runnable aren’t sovereignty: here the compute isn’t an obstacle — 462 MiB, a laptop CPU, no cluster to get approved. And that’s exactly why the discipline of the register matters more, not less: once hardware stops being a barrier, anyone can do the download, with no procurement department any the wiser. The same question we ask about an 800 GB model applies to a 462 MiB one: who downloaded it, under which licence archived in the text of the day, with which approved uses and approved by whom. The technical report a model card ought to provide before installation is nearly all present here: architecture, hardware, evaluation method, limits openly admitted. For a body that handles meeting transcripts — in government a dictated minute is as official a document as the act that follows it — the naming clause changes nothing for internal use: it’s downstream distribution, if the model ends up inside a resold product, that belongs in the register alongside version and digest.
What we don’t know
We haven’t run the model: no verification of ours on the declared accuracy (94.8%), calculated by the publisher on its own test set, not included in the repository. We don’t know what transcripts it was trained on, nor whether they contain real voices collected with verifiable consent. We don’t know the organisation’s corporate structure: no registered office, no legal form stated. We make no legal qualification of the additional licence clause, a text for whoever has to apply it to read, not an opinion of ours. We report the state verified on 22 August 2026; a link that works today may stop working tomorrow, like the v1 tag that today, simply, doesn’t exist.
See the service · Talk to an engineer
The register, applied to this case
Complying: for a model like this, the register can’t stop at marking open licence, all in order. It has to archive the additional clause in full — the name to be kept in every distribution — with the date it was read, and record that the badge on the tagged release is a broken link, so it isn’t repeated tomorrow as if verified today. It’s the record that holds up in front of an inspector, a client in a tender, a board.
Deciding: the same system brings models, data, contracts, archives and documents together into a single operating model, on which AI agents execute decisions with a human operator in command, for large enterprises, defence, government and healthcare. A 462 MiB model on a laptop CPU doesn’t need a cluster to enter the perimeter — it needs someone to register it anyway, with the same discipline as an 800 GB one: it’s the volume of small models, downloaded with nobody the wiser, that is the real risk once hardware stops being an obstacle. Always in two modes, on-premise on self-contained machines or dedicated cloud with a dedicated VPN and a data centre in Italy, always with shared management.
From the first session, at no cost, comes the dated register of the models you have in service — including the small ones, downloaded with procurement none the wiser: for each one, the licence archived in the text of the day it was downloaded, not a badge promising a tag that doesn’t exist today. It stays yours even if we don’t go on to work together.