SAT, SEPTEMBER 05, 2026
Independent · In‑Depth · Practitioner‑Tested
✎ Large Language Models

You Cannot Audit a Weights File. You Can Audit This One

The Institute of Foundation Models released K2 Horizon — six models from 0.9 billion to 375 billion parameters, published with weights, training code, training data, checkpoints and training logs. Every other open-weight release this summer shipped the weights and kept the rest, which is precisely the gap the Open Source Initiative has been objecting to.

By AIToolsRecap September 5, 2026 6 min read 25 views
Home Articles Large Language Models Somebody Finally Released the Training Data — S...
WHAT WAS RELEASED

● Six models, 0.9 billion to 375 billion parameters, from the Institute of Foundation Models.

● Included: weights, training code, training data, checkpoints and training logs.

● Why that is different: every other open-weight release this summer shipped weights and withheld the rest.

The gap this closes

Yesterday we wrote about the licensing drift in open-weight releases — Qwen 3.8-Max under custom terms rather than Apache, Kimi K3 under a modified MIT, GLM-5.3's flagship terms still unstated. The Open Source Initiative's objection to all of it is that MIT and Apache were written for source code you can read, and a weights file is not that.

You cannot inspect a weights file, reproduce it without the training data, or verify what went into it. A licence granting freedom to modify something you cannot examine grants less than the same words grant for source.

TRAINING DATA IS THE PART NOBODY PUBLISHES

Weights are the artefact. Training data is the thing that would let someone else reproduce, audit, or challenge what you built.

It is also the part that carries legal exposure — a federal court has already held that acquiring training material through piracy is unlawful, and Sony and Warner sued Anthropic last week on exactly that question. Publishing your corpus is an act of confidence as much as openness.

The range is the other notable part

Six models from 0.9 billion to 375 billion parameters is a spread, not a flagship. The small end runs on a laptop; the large end needs a serious node.

That matters for research more than for production. A family trained the same way at different scales is what lets people study how behaviour changes with size — which is hard to do when every lab publishes one model and keeps the ladder private.

Combined with the checkpoints and logs, it is the most complete public record of how a model family was actually built that anyone has released this year.

What to check before using it

Check Why
The licence file Publishing everything is not the same as permitting everything. Read the terms
What the training data contains Now that you can look, look. That option does not exist elsewhere
Benchmarks against the majors Full openness is a governance achievement, not a capability claim
Which size you actually need Six options means the smallest that works is probably not the largest

Worth being clear on the last one. Nothing in this release claims frontier performance, and it would be a mistake to read full disclosure as a benchmark result. What it offers is verifiability, and for some uses — regulated work, research, anything where provenance matters — that is worth more than a few points.

Sources

FAQ

What is K2 Horizon?

A family of six models from the Institute of Foundation Models, ranging from 0.9 billion to 375 billion parameters, released with weights, training code, training data, checkpoints and training logs.

Why is releasing training data significant?

It is the part that allows reproduction and audit, and the part that carries legal exposure. Almost no lab publishes it, which is why open-weight and open-source have drifted apart.

Does this make it better than Qwen or GLM?

Not on capability — nothing here claims frontier performance. It is more open, which is a different axis and matters most where provenance does.

Can I use it commercially?

Check the licence file. Publishing everything and permitting everything are separate decisions.

Which size should I use?

The smallest that does your job. Six sizes trained the same way is unusual and lets you test that properly rather than guessing.

Tags
Open SourceModel ReleaseLicensingTraining DataResearchHugging Face2026

Spot an inaccuracy?

We verify facts before publishing and correct errors promptly. If something in this article is wrong or outdated, let us know.

Report an error →