Every few days another lab announces an open-weight model, and most of us read that as good news and move on. It used to be a binary. You could download it and do what you wanted, or you could not.

That is not what the phrase means anymore, and the gap between the announcement and the license has gotten wide enough to cause real problems. Four major open-weight releases landed this summer. One was genuinely open. The other three each came with something the headline did not mention.

None of this is dishonest. These labs are shipping more openly than almost any American company, and the terms are published if you go looking. But the marketing has drifted well ahead of the reality, and if you are planning to build on any of it, the announcement is not the thing you should be reading.

Moonshot published the full weights for Kimi K3 in late July, a 2.8-trillion-parameter model and the largest open-weight release in history. Genuinely significant, and it sits on Hugging Face for anyone to download.

Everyone assumed it would carry the Modified MIT license Moonshot used for K2.7 Code. It did not. Hugging Face lists it as "other," and it ships under a bespoke document Moonshot wrote called the Kimi K3 License.

The terms are MIT-like with commercial conditions attached. Reporting on the license describes gates for anyone hosting K3 as a paid API or model-as-a-service above a revenue threshold, requiring a separate arrangement with Moonshot. The template it derives from also carried branding requirements above certain user and revenue counts.

For research, internal use, or a smaller product, that is effectively wide open. If you are building something that competes with Moonshot's own API, it is not. And a license that changes your rights based on your revenue is not open source by the standard definition, whatever the announcement says.

Alibaba's Qwen3.8-Max was the first Max-class Qwen ever released as open weights, published on Hugging Face and ModelScope in mid-August after a promise made on August 3 that had already slipped.

Two things about that release deserve attention.

The published weights are reportedly text-only and do not carry the one-million-token context the API version advertises. So the model you can download is not the model Alibaba sells you through QwenCloud, which means an evaluation you ran against the API does not transfer to the self-hosted version. It ships under a new revenue-share license rather than anything standard.

And Qwen3.8-27B, the small dense model promised in the same window, never arrived. No repository, no model card, no new date. That was the release most developers actually wanted, since community estimates put a Q4 quantization around 16 to 17GB, meaning it would run on a single 4090. A 2.4-trillion-parameter mixture of experts is a datacenter artifact. The 27B was the one you could own.

Zhipu released GLM-5.3 in mid-August with a striking claim: a 50 percent coding improvement over GLM-5.2 achieved entirely through extended post-training on an unchanged base model, plus first place among open-source models on Terminal Bench 3.0.

It was announced as the strongest open-weights coding model available. The weights were not available. Zhipu said they would follow in roughly two weeks pending security reviews, and did not disclose the license.

That is a reasonable thing to do, and the security review is a defensible reason for the delay. It is also a model marketed on its openness during a window when the only way to use it was Zhipu's own paid coding plan. Worth knowing which of those you are reading when the headline arrives.

Then there is the one that did it cleanly. Meta released Muse Glimmer under Apache 2.0, a 30-billion-parameter agentic model distilled from its flagship, with four-bit quantization bringing it to roughly 18 to 20GB so it fits on a 24GB consumer card.

Apache 2.0 has no revenue threshold, no bespoke terms, no branding requirement, and no commercial gate. You can build a business on it and never talk to Meta. The release also shipped complete: BF16 weights, a GGUF quantization, an ExecuTorch build, a small assistant model, and a speculative decoding drafter, with Ollama, LM Studio, vLLM, llama.cpp, and MLX support arriving in the same week.

Glimmer is not the most capable model in this issue. Meta says so itself, rating it below its own frontier threshold. But on the single dimension that determines whether you can actually build on something, it is the only one of the four with no asterisk.

The announcement will not tell you what you need to know. These four will.

Are the weights actually published? Not promised, not coming in two weeks. On Hugging Face with a working download today. GLM-5.3 was announced as the strongest open model on the market while the weights did not exist.

Is it the same model the API sells? Check the context window, the modalities, and the parameter count against the hosted version. Qwen's published weights reportedly drop the million-token context and the multimodal input. Any evaluation you ran on the API may not transfer.

What does the license actually say? Look at the license field on Hugging Face. If it says "other," someone wrote a custom document and you need to read it. Specifically look for revenue thresholds, user-count triggers, branding requirements, and restrictions on offering the model as a service. Apache 2.0, MIT, and standard Modified MIT are safe. Anything bespoke is not, until you have read it.

Can you actually run it? A 2.4-trillion-parameter model being downloadable is not the same as being usable. Check the accelerator requirement, whether your inference stack supports the architecture, and whether community quantizations exist. Kimi K3's attention mechanism was not supported in llama.cpp, Ollama, or LM Studio at release.

The stories moving fast beyond today's headlines:

Kimi K3. Weights published, 2.8 trillion parameters, largest open release ever. But it ships under a bespoke Kimi K3 License listed as "other" on Hugging Face, not the Modified MIT everyone assumed, with reported commercial gates above a revenue threshold for anyone hosting it as a paid service.

Qwen3.8-Max. First Max-class Qwen released openly, but the published weights are reportedly text-only without the million-token context the API advertises, under a new revenue-share license. The 27B checkpoint most developers wanted never shipped.

GLM-5.3. Announced as the strongest open-weights coding model, with weights promised roughly two weeks later pending security review and no license disclosed. Marketed on openness during a window when the only way to use it was Zhipu's paid plan.

Muse Glimmer. Meta's 30B agentic model under Apache 2.0, no revenue threshold, no bespoke terms, no gate. Complete release with quantizations and day-one support across Ollama, LM Studio, vLLM, llama.cpp, and MLX. Not the most capable model here, but the only one with no asterisk.

How to check. Are the weights actually downloadable today. Is it the same model the API sells. What does the license field say, and if it says "other," read the document. And can your stack actually run the architecture. The announcement will not answer any of those.

Stay building. 🤖