LFM2.5 from Liquid AI: above ten million dollars, the licence is not granted

Liquid AI has published the LFM2.5 family on Hugging Face, and its text model — LiquidAI/LFM2.5-2.6B, created on 28 July 2026 and updated on 7 August — carries, among the tags on its card, the word that matters most to readers of these notes: edge. Alongside lfm2, text-generation, liquid, lfm2.5, conversational, and the languages it covers — Arabic, Chinese, English, French — the card itself states what the model is built for. In eleven days it has drawn 77,973 downloads and 377 likes. The quantised variant, LiquidAI/LFM2.5-2.6B-GGUF, published on 1 August, is the one that actually runs on a small machine: another 31,489. A model for inference inside your own perimeter, with no network, is what large enterprises, government, defence and healthcare are looking for. Then you scroll the card to the bottom: license: other, license_name: lfm1.0. That is where the real work starts.

Three repositories, one licence

Under the same commercial name, Liquid AI keeps at least three separate repositories. LiquidAI/LFM2.5-2.6B: created 28 July 2026, eleven files, 77,973 downloads. LiquidAI/LFM2.5-2.6B-GGUF: created 1 August, 31,489 downloads. LiquidAI/LFM2.5-VL-450M, the family’s vision-language model: created 8 April 2026, 74,190 downloads, four commits in total, all on 8 April. The licence has never changed since publication: it was there from the first commit, not bolted on later to satisfy a reviewer — unlike the recent case of Qwen3.8-Max, whose weights arrived with the word “licence” still missing. All three declare the same line: license: other, license_name: lfm1.0.

A licence that looks like Apache 2.0

The document is titled the “LFM Open License v1.0”, licensor Liquid AI, Inc., 10,574 characters. Its structure follows Apache 2.0 almost section for section: Definitions, Grant of Copyright License, Grant of Patent License, Redistribution, Submission of Contributions, Trademarks, Disclaimer of Warranty, Limitation of Liability, Accepting Warranty. The redistribution conditions are ones anyone who has read an open-source licence knows by heart: a copy of the licence to recipients, prominent notices on modified files, the copyright, patent, trademark and attribution notices kept intact, the NOTICE file reproduced. Skim it for the familiar words and you find them all — as in other open-weight licences that look permissive at first glance. The problem is that there are two more sections, and Apache 2.0 does not have them.

The definitions that decide everything

Section 5, “Commercial Use Limitation”, reads: “(a) The rights granted under this License for Commercial Use are conditioned upon You or Your Legal Entity not exceeding the Threshold. (b) Any Commercial Use of the Work or a Derivative Work by a Legal Entity that exceeds the Threshold is not licensed under this Agreement. (c) The Threshold shall not apply to a Qualified Non-Profit Organization’s use of the Work or a Derivative Work for Non-Commercial or Research Purposes.” Rights for commercial use hold only below the threshold; above it, that use is not licensed; the exception applies only to a qualified non-profit, for non-commercial or research purposes. Section 2 grants the licence “including the Commercial Use limitation set forth in Section 5” — the limit is not a footnote, it sits inside the grant.

Three definitions, from Section 1, decide the rest. “‘Threshold’ shall mean annual revenue of 10 million United States dollars ($10,000,000) or more” — an annual revenue of ten million dollars or above. “‘Commercial Use’ shall mean any use of the Work for direct or indirect commercial advantage or monetary compensation” — broad, not confined to selling the model itself. And “‘Legal Entity’ shall mean the union of the acting entity and all other entities that control, are controlled by, or are under common control with that entity”, where control is triggered by fifty percent (50%) of the outstanding shares. What is measured is not the subsidiary: it is the group. The boundary here is not territory — as in a recent case — it is who you are, measured in revenue.

A question that needs asking, not answering

Does an Italian public authority running LFM2.5 for an institutional task count as “Commercial Use”? The definition is broad and does not exclude a public body as such. The only stated way out is letter (c), built on the “Qualified Non-Profit Organization”: an entity organised and operated exclusively for religious, charitable, scientific, public-safety testing, literary or educational purposes, exempt from federal income tax under Section 501(c)(3) of the United States Internal Revenue Code of 1986, “or any equivalent non-profit or charitable organization in a foreign jurisdiction”. Whether a municipality, a health trust or a ministry falls within that equivalence is not written in the licence: it is a legal judgement, to make and date before production, not after an inspection turns it up.

How we check it

When a client brings us an open-weight model, we do not stop at the Hugging Face card: we open the LICENSE file, isolate the clauses a commercial description never mentions — revenue thresholds, the definition of an entity, termination conditions — and put them in a register that stays with the client, not with us. For LFM2.5 that means calculating the ten-million threshold across the group, not the subsidiary, and dating the check. It is the same review we run, as an ongoing service, on every model an organisation puts into production.

See the service · Talk to an engineer

The section almost no one invokes, until it matters

The second section Apache 2.0 does not have is Section 11, “Termination”: “This License will terminate automatically and immediately if You fail to comply with any of its terms and conditions. Upon termination, You must cease all use of the Work and any Derivative Works and delete all copies in Your possession.” The licence ends automatically on the first breach, and from that moment the obligation is to stop all use and delete every copy. That is not theoretical when the model is built, by design, to run on the device: a system that cannot list where the copies of an artefact ended up cannot discharge that obligation.

The other side

This deserves the same rigour applied so far: a revenue threshold is not a deception. It sits in Section 5, in plain sight, and the text downloads in one click from the same card as the weights. A vendor that publishes a model for researchers, small businesses and non-profits, while keeping the direct relationship for anyone above ten million in revenue, is choosing a legitimate business model, stated openly. The problem is not the clause: it is that it gets read afterwards, once the model is already inside a project, and that the label these models circulate under — “open”, “open weights” — does not distinguish an unconditional permission from an invitation to negotiate.

What to do about it

Open the LICENSE file before downloading: the card’s label says only other.
Calculate the threshold across the group, not the single affiliated company: threshold 50% of the shares.
For every model in production, record: artefact, repository, revision, licence, particular conditions, date of verification.
Treat Section 11 as an operational risk: termination requires deleting the copies, and you need to know where they are.
Assess and date, before going into production, whether your organisation falls within the exception in letter (c).

How we solve this

Both axes apply here too. Comply: verifying the licence, provenance and threshold of every model in production becomes a check that runs on the client’s own systems, with a register — artefact, repository, revision, hash, licence, thresholds, date of verification — ready to show an inspector or a client. Decide: the same system brings the organisation’s models, data, documents and systems into a single operating model, on which AI agents execute decisions with a human operator in command, for large enterprises, defence, government and healthcare. The multi-model argument shows up concretely: if a licence excludes your organisation on revenue grounds, you swap the component — not the project — because the ontology, the data and the processes remain yours. Always in two delivery modes: on-premise on self-contained machines with no deep integration into the client’s network, or dedicated cloud with a data centre in Italy; always with shared governance.

Assessing LFM2.5, or another model built for the edge, and unsure whether your organisation stays under the threshold? Talk to one of our engineers: the first session is at no cost.