Z.ai releases GLM-5.3 as open-weight, claiming open-source SOTA on coding and cyber benchmarks

Z.ai has released GLM-5.3 as an open-weight model, with weights published under the zai-org organisation on Hugging Face. According to the model card, the release reuses the base model from GLM-5.2, meaning every claimed improvement comes from post-training rather than a new pre-training run. The gains Z.ai highlights are in complex coding and long-horizon agentic work.

What Z.ai claims

The company reports a 50% improvement over GLM-5.2 on its internal Z.ai Code Bench and calls GLM-5.3 the strongest open-weight model for coding, with best-in-class results among open-source models on public benchmarks including Terminal Bench 3.0 and Agents' Last Exam.

Selected results from the published comparison table:

• Terminal Bench 3.0: 28.3, up from 4.6 for GLM-5.2. Closed models still lead here, with Fable 5 at 33.7 and GPT-5.6 Sol at 34.6.
• Terminal Bench 2.1: 88.2 versus 81.0, level with Kimi K3 at 88.3 and just behind GPT-5.6 Sol at 88.8.
• DeepSWE v1.1: 66.9 versus 46.2, with Fable 5 at 69.7 and GPT-5.6 Sol at 72.7 still ahead.
• Agents' Last Exam: 28.5 versus 23.8, fractionally behind GPT-5.6 Sol at 28.6.
• SWE-Marathon v1.1: 42.5 versus 19.4, though Kimi K3 at 48.1 and Opus 4.8 at 48.8 score higher.

The pattern across the table: GLM-5.3 roughly doubles its predecessor on the harder, longer-running suites and closes much of the gap to closed frontier models without eliminating it.

Cyber capability emerged faster than expected

The most unusual disclosure concerns offensive security. Z.ai says that as it scaled post-training, cyber capability grew faster than it had anticipated. On CyberGym, which measures vulnerability discovery, GLM-5.3 posts 84.5 versus 77.2 for GLM-5.2, the top score in the table, ahead of GPT-5.6 Sol (83.6), Fable 5 (83.8) and Kimi K3 (80.0).

The largest gains sit further along the exploitation chain, according to the model card. ExploitBench jumps to 54.4 from 24.4, and ExploitGym reaches 105 points at a two-hour budget and 130 at six hours, against 29 and 39 for GLM-5.2. Closed models retain a wide lead on full exploitation: Fable 5 reports 181/247 and GPT-5.6 Sol 216/293 on ExploitGym, and 78.0 and 76.5 on ExploitBench.

Because the weights are downloadable, this capability ships without an API gate. The model card does not describe additional access restrictions or safety filtering tied to the cyber results.

Running it locally

The model card lists supported serving frameworks: SGLang, vLLM, TokenSpeed, Transformers, KTransformers and Unsloth. Ascend NPU hardware is supported through vLLM-Ascend, xLLM and SGLang.

Two configuration notes stand out. The reasoning_effort parameter accepts low, high and max, defaulting to max; Z.ai says to keep the default when reproducing leaderboard results. The chat template's clear_thinking option defaults to false, and Z.ai recommends explicitly passing clear_thinking=true for chat scenarios.

Caveats on the numbers

Almost every result is self-reported by Z.ai. Many evaluations ran inside the Claude Code harness, and the HLE evaluation used GPT-5.6-luna as a judge. Two rows come from outside parties: Artificial Analysis measured GDPval-AA v2, and Proximal ran FrontierSWE. The footnotes also disclose that Z.ai modified anti-cheat checks in PostTrainBench and SWE-Marathon, replacing pattern-matching checks it found too broad with LLM-based inspection.

Note the scope of the records as well: best among open-source models is not overall leadership. Fable 5 leads PostTrainBench at 41.8 versus 39.8, and closed models stay ahead on Terminal Bench 3.0, DeepSWE and the exploitation suites.

Why it matters

A downloadable model that competes with closed frontier systems on coding agents changes the calculus for teams that need self-hosting, data control or predictable inference costs. The release also shows how much headroom post-training alone now offers: the same base model delivers roughly doubled scores on the hardest suites. And the fast-emerging exploitation capability is precisely the kind of dual-use skill that makes open-weight releases a governance question, not just an engineering one. Until independent evaluations appear, the benchmark table is best read as Z.ai's own account of its model.