Z.ai, the Chinese AI lab formerly known as Zhipu AI, has released the full weights of its GLM-5.3 model on Hugging Face, turning one of the strongest coding and agentic models available into a free download. The release went live on August 28, just days after the lab confirmed that the viral stealth model known as "Ox Alpha" was in fact its own GLM-5.3-Flash, and it fulfills the open-weights promise that had the community watching closely all week.
The model page quickly collected more than 1,100 likes, and the Hacker News thread announcing the release climbed past 600 points as developers digested what is now sitting in a public repository: 141 safetensors files, shipped in FP8, comprising what may be the most capable openly downloadable coding model to date. For more context on this story, see our ongoing breaking AI news.
What Z.ai Actually Shipped
GLM-5.3 is a Mixture-of-Experts model built on an architecture Z.ai calls glm_moe_dsa. According to the configuration published alongside the weights, the network stacks 78 layers with 256 routed experts, activating 8 experts per token plus one shared expert. The context window spans 1,048,576 tokens — a full million — and the model supports a "reasoning_effort" parameter with low, high, and max settings, defaulting to max.
Perhaps the most interesting note in the model card is a candor about where the improvements came from: GLM-5.3 uses the same base model as GLM-5.2, and every gain comes from post-training. The lab pointedly describes the result as "the most capable open-weights model for coding."
Benchmarks: Open-Weights SOTA on Agentic Coding
The published evaluation table is extensive and unusually specific about its methodology, listing results against Kimi K3, DeepSeek-V4 Pro-0813, Qwen3.8-Max, Anthropic's Opus 4.8, Google's Fable 5, and OpenAI's GPT-5.6 Sol.
On Terminal Bench 3.0, GLM-5.3 scores 28.3 — open-source state of the art, and a jump from 4.6 for GLM-5.2, though still behind the closed frontier of Fable 5 at 33.7 and GPT-5.6 Sol at 34.6. On Terminal Bench 2.1 it posts 88.2, essentially level with Kimi K3 and GPT-5.6 Sol and ahead of Opus 4.8's 85.0. DeepSWE results jumped from 46.2 to 66.9 between versions, and GLM-5.3 tops the table on AutomationBench at 48.2 and on GDPval-AA v2 at 1769, the latter evaluated by Artificial Analysis.
Z.ai also reports a 50% improvement over GLM-5.2 on its in-house Code Bench, alongside open-weights-best results on benchmarks like Agents' Last Exam.
The Cyber Capability Question
The model card is remarkably blunt about a development that safety researchers have been tracking for months: "As we scaled post-training, cyber capability developed faster than we expected," the card states. GLM-5.3 is state of the art on CyberGym for vulnerability discovery at 84.5, edging out every model on the table including GPT-5.6 Sol, and it more than doubles GLM-5.2 on exploitation benchmarks — 54.4 versus 24.4 on ExploitBench, and 105/130 versus 29/39 on ExploitGym's two time budgets.
AI Buzz Wire covered these capabilities when they emerged with the API-only release in mid-August. What changes now is distribution: the weights are downloadable, quantized community builds are likely to follow quickly, and anyone with enough memory can run the model locally. In the Hacker News discussion, some users noted the model felt less restrictive around security-adjacent tasks than US-hosted alternatives — community impressions rather than an official policy claim, but reflective of why this release is being discussed in security circles as much as developer ones.
A License Aimed at Hyperscalers
Unlike GLM-5.3-Flash, which shipped under a clean MIT license, GLM-5.3 arrives with a custom "GLM-5.3 License." The New Stack characterized it as aiming at hyperscalers, and the text bears that out. For nearly everyone, the terms are effectively permissive: use, copy, modify, merge, publish, distribute, sublicense, and sell, with just a requirement to include the notice.
The exception targets one specific business model. If a licensee or its affiliates operate a "Model as a Service" business — offering third parties inference or fine-tuning access — and their combined revenue exceeds $10 billion over any consecutive 12 months, they must pass a Z.ai security review before any commercial use of the model or its derivatives. That clause draws a circle precisely around the biggest cloud providers, while leaving startups, researchers, enterprises, and mid-scale hosts untouched.
The two-track approach is telling. The Flash variant, now revealed as the Ox Alpha model that went viral before TechCrunch uncovered its origins, runs entirely on Chinese silicon and carries the MIT flag for maximal adoption. The flagship gets a bespoke license that keeps hyperscaler deployment — particularly by American clouds — gated behind Z.ai's own review process.
Running It Yourself
For those inclined to try, the model card lists deployment support across vLLM, SGLang, KTransformers, Unsloth, and Hugging Face Transformers, with cookbook recipes for several of them. Notably, Huawei Ascend NPU deployment is documented through vLLM-Ascend, xLLM, and SGLang — continuing the lab's push to prove frontier models run on non-Nvidia hardware. In the community, DeepInfra was first to offer a third-party endpoint of the open model on OpenRouter.
Z.ai also published a technical report on arXiv, titled "GLM-5: from Vibe Coding to Agentic Engineering," covering the model family's development.
Why It Matters
The release compresses the gap between the closed frontier and the open frontier to a rounding error on many agentic coding benchmarks — weeks, not years. Combined with Kimi K3 and DeepSeek-V4, it means developers outside the biggest labs now have credible alternatives they can own, audit, and fine-tune.
It also signals a shift in how open releases are governed. The MIT-vs-custom-license split within a single model family suggests Chinese labs are learning to use licensing as strategic infrastructure: openness where it builds an ecosystem, leverage where it protects position. The next test will be whether any $10-billion-plus MaaS operator actually submits to a Z.ai security review — or whether the clause simply keeps the biggest clouds out while everyone else builds on top.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
