GLM-5.3-Flash open weights: what the model card says, and what it means for a 12GB card

What was released

Z.ai published the weights for GLM-5.3-Flash on 26 August 2026 under the MIT license. Every figure in the table below was read from the model card and the config.json in the repository (GLM53F-HF-001), not from reporting about the release.

The MIT license is the part with the fewest asterisks on it. Commercial use, modification and redistribution are all permitted, with no revenue threshold and no acceptable use annex bolted on. Several other open weight releases attach one or both, so this is worth stating plainly rather than filing under “open”.

One number in the model card needs disambiguating. A footnote next to one benchmark result mentions a maximum context length of 300,000 tokens. That describes how that specific evaluation was run, not the model. The architecture value is 1,048,576 (GLM53F-HF-001).

Multimodal in the model, not around it

Image and video are handled inside the model rather than by an adapter bolted to a text model. The config carries a vision configuration and dedicated image and video token ids (GLM53F-HF-001), which is what “natively multimodal” means here and is the difference that matters for anyone planning to quantize it: a bolted-on vision encoder can be dropped or kept separately, and an integrated one cannot.

It had already been running under another name

Before the release, the model was served anonymously on OpenRouter as “Ox Alpha”. OpenRouter’s own listing for GLM-5.3-Flash states that the stealth model was revealed to be this one (OPENROUTER-GLM53F-001). Anyone who benchmarked Ox Alpha during that window was benchmarking this release.

What running it locally would take

Z.ai documents local serving through SGLang, vLLM, TokenSpeed, Transformers, KTransformers and Unsloth (GLM53F-HF-001). So the software path exists. The memory question is separate, and it is the one that decides whether any of that is reachable on a desktop.

A mixture of experts model activates 18B parameters per token, and it is tempting to read that as the memory requirement. It is not. Activating a subset lowers the compute per token; the weights still have to be somewhere the runtime can reach quickly. At four bits per parameter, 320 billion parameters is roughly 160 GB before any KV cache or activation memory. That figure is arithmetic from the parameter count in GLM53F-HF-001 rather than a measurement, and the real number will be higher once the cache for a long context is added.

Against that, a 12GB card is not short by a margin that a setting can close. Neither is a 24GB one. The configurations worth measuring on hardware this size are the ones that accept the weights living in system RAM or on disk and pay for it in tokens per second, which is a different question from “does it fit” and needs its own numbers (GLM53F-HF-001).

So the honest answer today is that this site has not run it, and the interesting result is not whether it loads but what the throughput collapses to when most of the model is not on the GPU.

Where these numbers come from

The model card and config.json were read directly from the repository rather than through a rendered page (GLM53F-HF-001). Z.ai’s launch post (GLM53F-BLOG-001) is linked from that model card, but its text is rendered by a client side application and could not be extracted on retrieval, so nothing above rests on it. The developer documentation (GLM53F-DOCS-001) was retrieved and covers the model as a vision language model without carrying the license or the serving list.

What gets measured next

The next article on this model is a measurement, and it will report one thing: what a 12GB card and an M4 Pro do with GLM-5.3-Flash when the weights cannot be resident, with the quantization, the context length and the offload configuration stated for each run. If the answer turns out to be that neither machine reaches a usable rate, that result gets published too, together with the point at which it stopped being usable.