Z.ai ships GLM-5.3 weights for agentic coding

Z.ai has put GLM-5.3 on Hugging Face. That is the story.

On 28 August 2026 the lab’s official account posted that GLM-5.3 “is now open-weight,” calling it their most capable model for agentic coding and cyber defense and pointing practitioners to downloadable checkpoints plus the 14 August tech blog. The Hugging Face repo is live at zai-org/GLM-5.3: 753 billion parameters, safetensors in BF16 and other tensor types, a chat template, and local-serve notes.

The model itself is not a surprise. Z.ai launched GLM-5.3 on 14 August and said then that weights would follow in two weeks, after safety evaluation and hardening. What changed is access. You can pull the checkpoints instead of going through Z.ai’s API.

Same base, different post-training

Z.ai is blunt about the recipe. GLM-5.3 uses the same base model as GLM-5.2. “Every gain comes from post-training,” both the tech blog and the model card repeat. The company describes a month of scaling on the stack it already had for 5.2 — more environments, more diverse long-horizon tasks, more compute — rather than a new pretrain.

The card lists the model at 753B parameters. Named serving paths include SGLang, vLLM, TokenSpeed, Transformers, KTransformers, Unsloth, and Ascend NPU stacks. Thinking is not optional. GLM-5.3 exposes a thinking budget through reasoning_effort (low, high, or max), defaulting to max if the parameter is omitted. Disabling thinking is no longer supported. For chat, the card says to pass clear_thinking=true explicitly, because it defaults to false.

The Hugging Face model tree already shows finetunes and a set of quantized derivatives, including listings for llama.cpp, LM Studio, Jan, and Ollama. Inference providers named on the card include Together AI, Baseten, and Zai.

Vendor numbers, not a new leaderboard drop

Z.ai positions GLM-5.3 as a coding and long-horizon agent and publishes the same comparison table on the blog and the model card. Those figures are the lab’s own evaluations unless a footnote says otherwise. Treat them that way.

On coding, Z.ai reports Terminal Bench 3.0 moving from 4.6 (GLM-5.2) to 28.3, and DeepSWE v1.1 from 46.2 to 66.9. It also claims a 50% improvement over GLM-5.2 on an in-house Z.ai Code Bench, and open-source state of the art on Terminal Bench 3.0 and Agents’ Last Exam among the open-weight models it compared. Other vendor-reported coding scores on the same table include Terminal Bench 2.1 at 88.2, NL2Repo at 58.0, ProgramBench (Almost Solved) at 19.0, FrontierSWE at 78.1, SWE-Marathon v1.1 at 42.5, and PostTrainBench at 39.8.

The table is not a clean sweep against closed models. In Z.ai’s own numbers, GPT-5.6 Sol and Fable 5 (with fallback) sit ahead of GLM-5.3 on several coding rows, including Terminal Bench 3.0 (34.6 and 33.7 versus 28.3) and DeepSWE v1.1 (72.7 and 69.7 versus 66.9).

On the private Z.ai Code Bench, the lab says GLM-5.3 is both stronger and cheaper in tokens than 5.2 at every effort level. At Max effort it reports 34.5% end-to-end completion at roughly 75K output tokens per task, versus 23.4% at 96K for GLM-5.2. At High effort it reports 31.4% at around 50K output tokens, which it says beats Claude Opus 4.8 at 29.5% with 120K. It still trails Claude Fable 5 on that private set, listed at 39.5% at Max effort.

Agentic rows on the public table — still vendor-reported — include Toolathlon Verified at 73.0, AutomationBench v1.0.6 at 48.2, Agents’ Last Exam (ALE-CLI) at 28.5, HLE with tools at 62.5, and GDPval-AA v2 at 1769, the last of those attributed to Artificial Analysis.

Many of the coding and agent scores were run in a Claude Code harness with long contexts, high token caps, and multi-hour timeouts. Footnotes on the card spell out those settings in detail. The cited pages do not include independent third-party reruns.

Cyber claims, also vendor-reported

Z.ai also markets GLM-5.3 as a cyber-defense model and says cyber skill showed up faster than expected as post-training scaled. On its table, CyberGym is 84.5 versus 77.2 for GLM-5.2; ExploitGym is 105 / 130 tasks under 2-hour and 6-hour normalized budgets versus 29 / 39; ExploitBench is 54.4 versus 24.4. The lab says the gains are largest further up the exploitation chain, and that closed models still lead on those harder rows.

The blog further claims that, working with security teams in China, the models identified 2,436 vulnerabilities across 269 projects after expert review, including 1,097 medium-to-high severity issues, with a disclosure ledger tracking what is already public versus still under embargo. Those counts are Z.ai’s. They are not independently reproduced in the pages cited below.

That is a lab describing its own evals and disclosure program. It is not a playbook.

What you can actually run

The practical change is local serving. The card points at SGLang, vLLM, Transformers, and several other runtimes. The X post frames the drop as download, run, and customize — not another API SKU.

If you already used GLM-5.3 through Z.ai, the blog’s API note still matters: thinking.type: "disabled" will fail. Switch thinking on and pick an effort level before you change model IDs. For coding, Z.ai recommends max.

The 14 August post also said you could try the hosted model in ZCode, Claude Code, OpenCode, and similar agents via Z.ai’s coding plan. That path is unchanged. The new path is the Hugging Face weights.

What is still unknown

The fetched model card and blog do not spell out a license name in the prose available here, so the exact terms are unknown from these sources. Read the repo license file before you ship anything.

Independent reproductions of Terminal-Bench 3.0, DeepSWE, and the cyber suite are also unknown. Z.ai’s footnotes are unusually detailed about harnesses, timeouts, and anti-cheat tweaks. That is useful, and it is still self-reported.

If you want the checkpoints, start at the Hugging Face card, skim the serve notes for your runtime, and keep the 14 August blog open for the training story and the eval footnotes. Then run your own tasks. The interesting question is not whether Z.ai’s table looks good. It is whether a 753B post-train of GLM-5.2 holds up in your agent loop once it is off their API.