Several organisations publish the trained weights of models that cost enormous sums to produce, and then let anyone download and run them. The reasoning is commercial rather than charitable, and it follows a pattern that repeats across the technology industry.
What releasing weights actually means
Weights are the numerical parameters produced by training. Publishing them lets anyone run the model on their own hardware, indefinitely, without contacting the publisher.
That is different from publishing a paper describing the method, and different again from offering the model through an interface. Weights are the working artefact.
Once released they cannot be recalled. Copies propagate to mirrors and hard drives within hours, and a later change of policy affects only future versions.
The training data, the training code and the tuning recipes usually stay private. What is shared is the output of the process, not the process.
So the release transfers capability without transferring the ability to reproduce or improve it independently.
The commoditisation argument
A company whose revenue comes from somewhere else benefits when a rival's core product becomes free.
If a competitor sells access to models, and comparable models are freely available, that competitor's pricing power erodes while the publisher's own business is untouched.
The tactic is old. Firms have long funded free versions of the layer their rivals charge for, so that value concentrates in the layer they own.
For cloud providers, the layer they own is compute. A world full of freely available models is a world where more people rent servers.
The release is an investment in demand for the thing that is not being given away.
How it recruits and retains researchers
Machine learning researchers are scarce, mobile and publication-oriented. Many will not join a team whose work disappears into an internal system.
Permission to publish is a real term of employment, comparable to compensation, and organisations that withdraw it lose people.
A public release also creates external evidence of a team's capability, which matters for hiring the next cohort.
The effect compounds. A group known for shipping open artefacts receives stronger applications, which produces better artefacts.
Some releases are best understood as a hiring cost that happens to produce a public good.
Why distribution beats secrecy for a challenger
A newer entrant faces an attention problem before it faces a revenue problem. Developers build on whatever they already have installed.
Giving weights away buys adoption directly. Tutorials, fine-tunes, integrations and tooling accumulate around the model, and that ecosystem is expensive for a rival to replicate.
Every downstream project that depends on a particular model family raises the cost of switching away from it later.
The publisher can then monetise the position through hosted versions, support contracts, enterprise features or a larger closed model at the top of the range.
The free model is the funnel rather than the product.
What the release does not include
Running a large model is not free even when the weights are. Hardware, memory and engineering time are substantial, and they recur.
Serving a model reliably to many users is a separate discipline from having the model, which is why hosted versions of open models sell well.
Documentation is often thin. The reference implementation runs, but adapting it to a production system is work the publisher has not done for you.
Safety tuning travels with the weights and can be removed by anyone with modest resources, which is a known consequence rather than an oversight.
The gap between a downloadable model and a working service is where most of the actual cost sits.
How licences restrict what looks free
Many published models carry custom licences rather than established open source terms, and the difference is substantive.
Common restrictions include caps on the size of business permitted to use the model commercially, prohibitions on using outputs to train competing models, and named categories of forbidden application.
These conditions make the release something other than open source in the traditional sense, which is why the phrase open weights has largely replaced open source in careful writing.
Enforcement is uneven, but the terms still shape corporate adoption, because legal teams read them before deployment.
The licence is where the strategy becomes explicit: the release is generous towards the users who help and restrictive towards the users who compete.
Where the strategy breaks down
Releasing weights forfeits usage data. A closed provider sees what people ask, which failures recur and where behaviour needs work; a publisher of open weights sees none of that.
Training runs are expensive, and the funding case weakens if the resulting model is immediately free to competitors.
Regulatory pressure adds friction, since obligations that assume a controllable deployment sit awkwardly on an artefact that has already been copied worldwide.
Some organisations have narrowed their releases over time for these reasons, publishing smaller models while keeping the largest ones private.
The pattern is a spectrum rather than a binary, and firms move along it as their position changes.
Why a leader releases older models instead
An established provider has the opposite incentive to a challenger, and usually releases a previous generation rather than the current one.
The older model has already earned its revenue, and releasing it undercuts challengers whose best offering sits at roughly that capability level.
It also generates goodwill and research activity without exposing the frontier system that differentiates the paid product.
The timing follows the internal roadmap, which is why open releases from large providers tend to arrive shortly after a new flagship launch.
Reading a weight release as a statement of values misses most of what is happening; reading it as a position in a market explains the timing, the licence and the choice of which model went out.