GLM-5.3 open weights land, under a bespoke licence rather than MIT

When we covered GLM-5.3’s launch on 14 August, the model was a hosted product only, the weights were promised in “roughly two weeks” after a security review, and we said we would cover it properly and add it to the catalogue once they actually landed. They have now landed. Z.ai published the open weights in the final days of August 2026, on Hugging Face as zai-org/GLM-5.3, with a separate BF16 variant alongside. The promised timeline held, which is worth noting given how often “weights coming soon” quietly slips.

What was released

The full GLM-5.3 is a server-class mixture-of-experts of roughly 744 billion parameters, with about 40 billion active per token (the model card reports a total nearer 753 billion). It shares its base model with GLM-5.2, and Z.ai credits the improvement to roughly a further month of reinforcement learning on executable environments and longer tasks rather than a new pretraining run. It carries a 1-million-token context window and up to 128k output tokens, and it is a reasoning model with always-on thinking: it exposes three effort levels, and unlike earlier versions it no longer lets you turn thinking off.

Alongside it, Z.ai shipped GLM-5.3-Flash, the first natively multimodal model in the GLM-5 line. Flash is a smaller mixture-of-experts, reported at 320 billion parameters with about 18 billion active, handling text, images, video, and visual documents, and it also carries a 1-million-token context. The two are licensed differently, and that difference is the part worth reading carefully.

The licence: not the MIT release we thought was plausible

In August we wrote that GLM-5.2 had been MIT, so a permissive release for 5.3 was plausible, but that we would confirm rather than assume. Confirming was the right call, because the answer is more nuanced than a simple yes.

The full GLM-5.3 is released under a bespoke GLM-5.3 License, not MIT. Read at source, it permits commercial use, and carries one real condition: if a licensee runs a Model-as-a-Service business and the aggregate revenue of the licensee and its affiliates exceeds 10 billion US dollars over any consecutive 12 months, it must pass Z.ai’s security review before using the model commercially. There is no geographic or territory restriction in the text. For the overwhelming majority of users, in other words, it is freely usable commercially, and the gate only reaches hyperscale model-hosting providers. GLM-5.3-Flash, by contrast, is released under the standard MIT licence, with no such condition. As always, read the licence text yourself before relying on it commercially: we are describing it, not giving legal advice.

Why it matters, and who it is for

The headline framing at launch was coding, and Z.ai positions GLM-5.3 as its strongest agentic coding model. That audience should still treat the numbers with the caution we set out in August: the striking coding figures, including Terminal-Bench 3.0 rising from 4.6 to 28.3 and DeepSWE v1.1 from 46.2 to 66.9, are Z.ai’s own (Z.ai, August 2026), and independent third-party evaluations of the open weights remain limited. The figures are a starting point, not a verdict.

For local use, be clear-eyed about scale. A 744-billion-parameter mixture-of-experts is firmly server-class. Even at a 4-bit quantisation, a model this size needs on the order of 350GB or more of memory for the weights alone, which means a multi-GPU server or a high-memory machine, not a consumer graphics card. That is an estimate, and the exact figure depends on the quantisation and the serving stack. Production GPU servers such as vLLM and SGLang are the realistic route to running it. GLM-5.3-Flash, at roughly 320 billion parameters with 18 billion active, is still large but more tractable, and its MIT licence and native multimodality arguably make it the more broadly useful of the two for anyone who is not specifically chasing the top of the coding benchmarks.

Where it sits

This follows a run of open-weight releases we have tracked through the year, and it mirrors a pattern we saw only weeks ago with Qwen3.8’s open weights: a hosted launch first, then the downloadable weights and the real licence once a review completes. It also continues a quiet but important trend in these releases, which is that “open” is not one thing. GLM-5.2 was MIT, GLM-5.3-Flash is MIT, and the flagship GLM-5.3 is a bespoke licence with a revenue-gated condition. The direction of travel is still towards capable open models arriving quickly, but the terms attached to them are worth checking on each release rather than assumed from the last.

We have added GLM-5.3 and GLM-5.3-Flash to the model catalogue with their confirmed licences and hardware requirements, and we will fill in independent benchmark results as they become available.