AI Foundationspredict · compress · act
act XI

The Open-Weight World

A model you can download is not the same as a model you can rebuild. The word 'open' hides a ladder — and almost everything sits on the bottom rung.

44

What 'Open' Means

decadeas of 2026-09
before this →The Open Question

For most of this guide, models were something you read about. Then they became files you could download. That change — the spread of open weights — moved who can build, who pays, and who is allowed to check the work. But “open” is used loosely, so let us be precise before we look at who is doing it.

1 · the weightsthe numbers you run and fine-tune — nearly universal2 · architecture & loading codehow to build the shape and load it3 · training codethe loop that made them4 · the training datarare5 · recipe, logs, and seeds — reproducible by a strangeralmost every ‘open’ model today sits on rung 1. a handful reach rung 4 or 5.
the openness ladder — each rung is a real, separate claim. Most 'open' models stop on rung 1.

The ladder

01
Weights — the parameters. You can run the model and fine-tune it. This is what 'open weights' almost always means.
02
Architecture — the shape of the network and the code to load it. Nearly always released, because otherwise the weights are unusable.
03
Training code — the loop, losses, and hyper-parameters that produced the weights. Sometimes released.
04
Data — the corpus the model learned from. Rarely released — it is the hardest part to hand over.
05
Recipe — data, code, logs and seeds together, so an outsider can reproduce the run. Very rare — OLMo and a few others.

Weights are not source

In software, open source means you can rebuild the program from its source code. For a model, the “source” is not the weights — it is the data and the recipe that produced them. Weights let you use a model. They do not let you rebuild it, audit what it was trained on, or verify a safety claim about it.

That mismatch is the whole argument. Software licences were written for code, where handing over the source is enough. A model has no such natural boundary, so the community spent years arguing about where the line should be — an argument the Open Source Initiative tried to settle in 2024 with its Open Source AI Definition, and did not fully settle.

Why share the weights at all?

four different motives, often mixed

Distribution. A free download earns users, talent, and mindshare faster than any marketing. Commoditisation. Giving away a capability removes a rival’s ability to charge for it. Scrutiny. Researchers can only study what they can run. Sovereignty. A country or company that cannot buy a closed API, or does not trust one, can host open weights itself.

the disciplineAsk which rung a model is on, and who benefits from calling it open. The two answers rarely match.
DOWNLOADyou have the numbers
RUNyou can use and fine-tune it
REBUILDyou still cannot — no data, no recipe

What you can and cannot do

With weights on rung 1, you can run the model, fine-tune it, measure its behaviour, and distil it into something smaller. You cannot know what it was trained on, verify the safety testing behind it, or be certain that the licence lets you do what you think it does — which is the subject of a later chapter. The rest of this act is a tour of the rungs: who climbed past rung 2, what they built, and what it cost.

introduces →open weightsopen-source AIopenness ladderpermissive licence
← previousThe Open Questionnext →DeepSeek