Counterfoil

,
Received into the cataloguecounterfoilSigned by D. Knauss

Local language models

Provisional review Last reviewed

Not fully scored yet. This may change. How we evaluate tools

In the convivial range. Of the six balances: 2 tipping, 2 holding, 2 not assessed. Position somewhere between 57 and 85 of 100. The balances not assessed could put it past the second watershed.

  • Biological degradation Not assessed Does this tool damage the living world, or the body, faster than either can recover? Not assessed: nobody has looked at this yet, so it is not counted as holding.
  • Radical monopoly Holding Can people still meet the need another way, without this tool? Holding: the tool does not cause this problem.
  • Overprogramming Tipping Does using this tool require training or certification that people cannot give each other? Tipping: the tool has begun to cause this problem, without going past the limit.
  • Polarization Tipping Does this tool move power or advantage to a few, at others’ cost? Tipping: the tool has begun to cause this problem, without going past the limit.
  • Obsolescence Holding Does this tool force people to replace what still works? Holding: the tool does not cause this problem.
  • Frustration Not assessed Does this tool give back less than it takes, counting everything it takes? Not assessed: nobody has looked at this yet, so it is not counted as holding.

What it does

Produces text in response to text, on a computer you control, with no connection to its maker. It works by predicting likely next words from patterns found in a very large body of writing.

Where it stops

At your own machine. It cannot report what you ask, and its maker cannot change or withdraw it. But it answers any question in fluent prose whether or not it is right, and it can only be made by the few organizations able to train one.

How easily you can stop using it

Delete it. Nothing else depends on it, and you can still read, write, and look things up yourself.

A disclosure first: a large language model drafted this record, at the editor’s direction. Read the scoring with that in mind.

This record is about language models that a person downloads and runs on their own computer. The same kind of software offered as a service by a company has its own record: Hosted language models. The two were one record until October 2026. They were split because the scores that put that record past the second watershed, above all obsolescence, describe the hosted service and not a copy you hold.

What is scored

  • The unit: A language model whose weights have been published, downloaded, and run with free software on hardware the user controls. Left out: hosted services, and the training of the model, which is done by its maker.
  • The job it is tested against: Producing fluent, relevant text on request. It is not tested here on whether the text is true; see the open questions. Models small enough to run on an ordinary computer are less capable than the largest hosted ones.
  • Where and when: The state of things in 2023 to 2025, from the makers’ own announcements and the Open Source Initiative. Nothing here has been measured by us.

What you get

Some makers publish a model’s weights, the numbers that make it work. Free software such as llama.cpp runs them on a personal computer with no connection to anyone. When Mistral AI released its first model in September 2023, it said it could be used “without restrictions,” including locally.

A model held this way is closer to a book than to a service. It cannot report what you ask it. Its maker cannot change it or take it back. It will give the same kind of answers in ten years, if you keep a machine that can run it.

What you do not get

You cannot make one. Training a model takes computing power that only a few organizations have, so the copy on your computer still comes from them. You can adjust it; you cannot rebuild it or inspect what it was built from.

Nor are all published models free to use as you like. The Open Source Initiative’s definition of open source AI (version 1.0, October 2024) asks for more than published weights, and the Initiative wrote in February 2025 that Meta’s Llama licence restricts who may use the model and for what.

And it is still a language model. What the record for the hosted kind says about fluent text that may not be true, and about a tool that writes in place of the person, applies here too.

A third arrangement: models held in common

Between a company’s service and a copy on one person’s computer there is a third arrangement: a model built or run by a public body, a co-operative, or a community, for itself. In September 2025 researchers at EPFL, ETH Zurich, and the Swiss National Supercomputing Centre released Apertus, a model whose architecture, weights, training data, and methods are all open and documented. It comes in two sizes, the smaller meant for individual use.

In Canada the sharpest form of the argument comes from First Nations. The First Nations principles of OCAP (ownership, control, access, and possession) were written for research data, and assert that a nation owns and controls information about itself. Writing in Policy Options in 2025, Margaret Yun-Pu Tu argued that language models built on Indigenous languages and knowledge taken without consent repeat an old pattern of dispossession, and that OCAP has hardly entered Canada’s policy on artificial intelligence.

Some of the work has started from the other end. At Mila in Montreal, the First Languages AI Reality initiative, co-founded by Michael Running Wolf, builds speech recognition for endangered languages in systems “explicitly designed to respect data sovereignty and linguistic self-determination.” BetaKit reported in July 2026 on Indigenous technologists building tools that keep data in the community that made it.

This matters for the scoring. Who holds a model, and on whose terms, is what the polarization balance asks. A model a community holds for itself may answer it differently from a copy one person downloads, and differently again from a company’s service. It does not answer the first question, whether the text is true. This record scores a copy held by one person. A model held in common is a different unit and may deserve its own record.

The six balances

Scored as set out in How we evaluate tools. Four of the six balances are assessed. It is scored as doing its job on the same decision as the hosted record; we have not measured how reliably a model of this size does it.

  • Biological degradation: Not assessed. Running a model at home uses a computer’s ordinary power. Training it used a data centre, and that cost belongs to the maker; how to share it out across copies is an open question.
  • Radical monopoly: 0. Nobody is obliged to run one. The pressure to use language models at all comes from their being built into other tools, and belongs to Hosted language models.
  • Overprogramming: 1. A tool that writes for people can take the place of the work of writing, wherever it runs. Setting one up also takes technical skill most people do not have. The evidence on what this does to learning is new and contested.
  • Polarization: 1. The weights come from the few organizations able to train them, and some licences limit their use.1 But a copy, once held, cannot be recalled, and anyone may pass on the freely licensed ones.
  • Obsolescence: 0. A downloaded model keeps working as it did. Nothing retires it but the loss of a machine able to run it.
  • Frustration: Not assessed. The cost of checking every answer is the same as for the hosted kind.

Where the threshold falls

In the convivial range, at 57: two of the six balances are tipping, so its position is 50 + 7. Two are not assessed, so this is the lowest the position can be, and the record is provisional. It rests on the same decision as the hosted record, that producing fluent text counts as doing the job.

Open questions

  • Does a language model reliably do its job? If the job is true statements, both records may belong below the first watershed.
  • How should the cost of training be counted against a copy that anyone may run? The same question arises for any tool that is costly to make and cheap to copy.
  • How capable are the models an ordinary computer can run, measured against the tasks people use hosted ones for? Find tests, with sources.
  • Does anyone depend on these who could not otherwise do the work? Score frustration from evidence, not from argument.
  • If most people only ever meet the hosted kind, is this record describing a tool or a possibility?
  • Catalogue a model held in common: a public model such as Apertus, or a language model a First Nation holds under OCAP. Who trains it, who holds it, who can leave, and who decides what it is used for?
  • Co-operatives own and run other shared tools. Is anyone running a language model as one? Find cases.
  1. Jordan Maris, “Meta’s LLaMa license is still not Open Source,” Open Source Initiative, February 18, 2025, https://opensource.org/blog/metas-llama-license-is-still-not-open-source. ↩︎

Sources

Review history

  1. Split from Large language models (now Hosted language models). Four of six balances assessed.
  2. Added a section on models held in common: public models, and First Nations data sovereignty under OCAP. No score changed.

Record last updated

Related tools

Part of a system:

  • Catalogue · Communication

    The smartphone

    Established

    Puts a telephone, a camera, a map, a wallet, and a general-purpose computer in a pocket, always connected, with its software supplied through its maker’s store.

    No. 020

  • Catalogue · Learning

    Vernacular literacy

    Established

    Lets people keep words outside memory: write them down, send them across distance and time, and read what others have written, in the tongues they already speak and for uses of their own.

    No. 018

  • Catalogue · Communication

    The fediverse

    Emerging

    Lets people on thousands of separately run servers follow and reply to each other, using a shared open protocol. Posts arrive in the order they were written; no company ranks them.

    No. 017