What are open weights?
An open-weight model is one whose trained parameters, the numbers produced by training, are published as files anyone can download. You can run it on your own machine, inspect it, fine-tune it and ship it inside a product without ever calling the original lab.
That is the whole of the definition, and it is narrower than most people assume.
What open weights are not
Open weights are not open source. In software, open source means you get the source code and a license that lets you read, change and redistribute it. The equivalent for a model would be the training data, the training code and the recipe that turned one into the other. Almost no major release includes those.
What you get is the finished artifact. You can see every parameter and still not know what the model was trained on, what was filtered out, or how the safety training was done. Reproducing the model from scratch stays impossible, which means the lab keeps the thing that is actually expensive.
Open weights are also not a single license. Terms vary from near-permissive to heavily conditioned, with clauses on commercial thresholds, acceptable use, and training competing models. Two models can both be called open and give you very different rights.
Why anyone prefers them
The reason is control, and it shows up in four places.
Cost stops being per-token. Once the file is on your hardware, inference costs electricity rather than a metered API bill, which matters for high-volume or long-running workloads.
Data stops leaving. For legal, medical and defense work, the deciding factor is often that nothing goes to a third party at all.
Availability stops being someone else’s decision. An API can be deprecated, rate-limited or repriced. OpenAI cut GPT-5.6 Sol pricing by over 20 percent in August 2026 and reinstated a five-hour usage limit for ChatGPT Plus subscribers days later. Both directions are outside a customer’s control. A downloaded model behaves the same next year.
Capability stops depending on model size alone. The XDA case is the useful data point: Qwen 3.8 27B, running locally, reverse-engineered a paid application’s license check in about 30 minutes, a job the writer had assumed needed a frontier model.
Why labs hesitate
A published weight file cannot be recalled. That single property drives most of the caution.
Safety training lives in the weights, and it can be removed. Anyone with the file and modest compute can fine-tune the refusals away, which means a release is effectively a decision about the model’s least-restricted version, not the one that ships.
Z.ai’s delay of GLM-5.3 in August 2026 is the clearest recent example of the calculus. The company found the model unusually good at finding and exploiting vulnerabilities, with 1,097 medium-to-high-severity flaws among 2,436 surfaced across 269 open-source projects, including in the Linux kernel. It pushed the weights back by two weeks rather than cancel them.
That is the shape of the tradeoff in practice. Delay buys time to fix what the model found and to prepare disclosure. It does not change the fact that after release the capability belongs to everyone.
How to read a release
Three questions settle most of it. What exactly was published, weights alone or weights plus data and code? What does the license permit for your specific use, not in general? And what hardware does it need at the quantization you can actually run?
The answers vary per release, and the word open in the announcement will not tell you any of them.
Quick answers
What is the difference between open weights and open source?
Open weights means the trained parameters are downloadable. Open source, in the sense the term carries for software, would also mean the training code and data are available under a license that permits reuse and modification. Most models called open are open-weight only: you get the finished model, not the recipe.
Can I use an open-weight model commercially?
It depends entirely on the license attached to that specific release. Some use permissive terms close to Apache 2.0. Others add restrictions such as user thresholds, acceptable-use clauses or bans on training competing models. The weights being downloadable says nothing about what you are allowed to do with them.
Why do labs delay releasing weights?
Because a release cannot be undone. Z.ai delayed GLM-5.3's weights by two weeks in August 2026 after its own testing showed the model surfacing 2,436 flaws across 269 open-source projects, 1,097 of them medium to high severity. Once weights are public, safety training can be removed by anyone with a modest amount of compute.
Do I need a data center to run one?
No. Smaller open-weight models run on a single workstation or a laptop with enough memory. In August 2026 a writer at XDA ran Qwen 3.8 27B locally on a Lenovo ThinkStation and had it work out how a commercial application verified its license in about 30 minutes, with no cloud involved.