On August 25, 2026, Z.ai uploaded the weights of its flagship model GLM-5.3 to Hugging Face. For the open-source community, this is the step that turns “open on a slide” into “actually downloadable”: a 753B-parameter model that was API-only at its August 14 launch can now be pulled and run locally. Three days later, “GLM-5.3 is now open-weight” reached the Hacker News front page with 806 points and 281 comments — the most-discussed open-weights release of the week.
From API-Only to Downloadable Weights
When GLM-5.3 launched on August 14, the official framing was “the strongest open-weights coding model,” but the weights themselves only landed on August 25. The model card restates two key points: GLM-5.3 reuses the GLM-5.2 base, with all gains coming from post-training, and the in-house Z.ai Code Bench score improved roughly 50% over GLM-5.2. The coding numbers are strong: Terminal Bench 2.1 at 88.2, FrontierSWE at 78.1, DeepSWE at 66.9, and 62.5 on HLE with tools. The security world is watching a different line: 84.5 on CyberGym for vulnerability discovery. Z.ai itself admitted at launch that the model’s cyber-attack capabilities were “developing faster than we expected” — an honest warning that feels more urgent now that anyone can download the weights.
753B MoE with Sparse Attention: The Hardware Bar
Let’s start with the conclusion: this is not a model you run on a consumer GPU. GLM-5.3 has 753 billion total parameters in a mixture-of-experts architecture with sparse attention. The checkpoint is roughly 756 GB across 141 safetensors shards, with most parameters stored in FP8 precision and a small remainder in BF16 and F32. Self-hosting means a multi-GPU server or a large-memory CPU inference box. The framework support list is long: SGLang, vLLM, Transformers, KTransformers, Unsloth, and TokenSpeed are all officially supported, alongside Huawei Ascend NPUs. If you would rather not touch hardware, six hosted inference providers are listed on the Hugging Face page: Z.ai, Together, Novita, Fireworks, DeepInfra, and Baseten. Day-one support from the standard serving stack matters, too: the same weights behind the hosted APIs can drop into a self-hosted SGLang or vLLM deployment without conversion.
A Custom License: Downloadable, But Read the Terms
Notably, the release does not use MIT. It ships under a custom “glm-5.3” license, tagged as “other” on Hugging Face. Compare that with August 26, when Zhipu open-sourced the much smaller GLM-5.3-Flash under MIT, and the strategy is clear: give away the small model to win the ecosystem and developer mindshare, keep control of the flagship. “Open weights” and “open source” are two different things here — anyone can download the weights, but the terms define the commercial boundary. Teams serious about adopting GLM-5.3 cannot skip the legal review.
What the Community Measured
The HN thread quickly became a large-scale hardware benchmark. Two DGX Sparks with NVFP4 quantization reached 20-30 tokens per second; a roughly $6,000 custom build managed about 14; an EPYC box with DDR4 and a few 3090s hit around 10; a decade-old dual-Xeon server crawled at about 1. Whether to buy a Mac Studio with 512 GB of unified memory, or pay anything at all to “own the whole inference stack,” sparked hundreds of comments — privacy, compliance, and vendor risk were the main arguments in favor. The one point of agreement: localizing a 753B flagship remains expensive, and the actual daily driver for most people will be a small model. Reception beyond the hardware talk was positive. One commenter called it “the sweet spot open weights model” — slightly less capable than Kimi K3 but easier to run, with expectations of better third-party pricing and speed. Others noted it is less restrictive than US models on security work, handling exploit analysis and reverse engineering where more guarded models refuse.
What It Means for Developers
Three things. First, model routing gains a self-hostable top-tier option: teams that need high-end coding capability but want to avoid closed APIs finally have a frontier open-weights fallback they can actually deploy. Second, the custom license is a compliance line — read the terms before adopting, and don’t be misled by the word “downloadable.” Third, the CyberGym numbers are a reminder that offensive capability in open weights is real; if you use this model for red-teaming or security research, budget for isolation.
Sources
- GLM-5.3 model card — Hugging Face
- GLM-5.3 announcement — Z.ai
- GLM-5.3 is now open-weight — Hacker News
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.