Zhipu AI releases GLM-5.3, claims strongest open-weights coding model

Zhipu AI has released GLM-5.3, an open-weights coding model built on the same base as GLM-5.2. Every reported improvement over its predecessor comes from extended post-training, with the company describing GLM-5.3 as the most capable open-weights model for coding and highlighting large gains in agent-based and vulnerability-discovery tasks.

What changed from GLM-5.2

According to Zhipu’s release notes, the underlying base model is unchanged from GLM-5.2. All reported gains came from scaling post-training using the stack Zhipu assembled for the previous release: IndexShare for long-context processing, SAO for reinforcement learning on long-horizon tasks, and slime for large-scale asynchronous training. Zhipu describes a month of further scaling that added more task environments, more diverse tasks, and more compute spent on them.

Coding benchmark results

On Zhipu’s in-house Z.ai Code Bench, GLM-5.3 shows a 50% improvement over GLM-5.2. On public benchmarks, GLM-5.3 reaches open-source state of the art on Terminal Bench 3.0 and Agents’ Last Exam.

Headline score changes from GLM-5.2 to GLM-5.3 include:

• Terminal Bench 3.0: 4.6 to 28.3
• DeepSWE v1.1: 46.2 to 66.9
• Agents’ Last Exam: 23.8 to 28.5
• Terminal Bench 2.1: 81.0 to 88.2
• NL2Repo: 48.9 to 58.0
• FrontierSWE: 67.5 to 78.1
• SWE-Marathon v1.1: 19.4 to 42.5
• PostTrainBench: 31.7 to 39.8
• ProgramBench Almost Solved: 9.5 to 19.0

On agentic benchmarks, GLM-5.3 also moves up: Toolathlon Verified rises from 59.9 to 73.0, and AutomationBench v1.0.6 climbs from 26.2 to 48.2. HLE with Tools moves from 54.7 to 62.5, and GDPval-AA v2 rises from 1508 to 1769.

How are the training environments built?

Zhipu says it shifted the training mix toward tasks that resemble real units of expert engineering and research work rather than coding exercises, with some tasks representing several days of work for an experienced engineer. In one example, the model is given the same working environment as an engineer, with access to compute clusters, storage systems, internal documentation, codebases, and experiment results, and must diagnose bottlenecks, implement optimizations, run experiments, and deliver a measurable end-to-end speedup.

The company describes a pipeline in which research agents collect task patterns from real work and convert them into runnable long-horizon environments with multi-step dependencies and hidden state, while a judge agent attempts each task to confirm it is solvable. Verifiers are synthesized without access to the reference solution, and solver trajectories are used to find and close reward shortcuts.

Token efficiency

At Max effort, GLM-5.3 reaches 34.5% on Z.ai Code Bench at roughly 75,000 output tokens per task, compared with 23.4% at 96,000 tokens for GLM-5.2. At High effort, GLM-5.3 reaches 31.4% at around 50,000 output tokens, surpassing Claude Opus 4.8 at 29.5% with 120,000 tokens. GLM-5.3 remains behind Claude Fable 5, which reaches 39.5% at Max effort.

Emergent cyber capability

Zhipu introduced vulnerability discovery data and environments into the training mix and reports that cyber capability developed faster than expected as training scaled. The company says GLM-5.3 began to reason across multiple stages of exploitation, forming coherent plans for complete exploitation chains.

On CyberGym, which starts from white-box source code and tests whether a model can identify and validate vulnerabilities by triggering faults, GLM-5.3 scores 84.5%, up from GLM-5.2’s 77.2% and ahead of Mythos 5 at 83.8% and GPT-5.6 Sol at 83.6%. On ExploitBench, which requires deeper reasoning about real vulnerabilities and exploitation, GLM-5.3 reaches 54.4%, more than doubling GLM-5.2’s 24.4%, while Mythos 5 and GPT-5.6 Sol score 78.0% and 76.5%. On ExploitGym, which measures exploitation tasks completed under time-normalized budgets, GLM-5.3 completes 105 tasks within two hours and 130 within six hours, compared with 29 and 39 for GLM-5.2; Mythos 5 remains ahead at 181 and 247 tasks.

The reported pattern is that the further up the exploitation chain a benchmark sits, the larger the gain over GLM-5.2 and the wider the remaining gap to the closed frontier.

Real-world vulnerability findings

Since GLM-5.2, Zhipu has worked with several security teams in China to run its models against real-world codebases. After expert review, screening, and deduplication, the model identified 2,436 vulnerabilities across 269 open-source projects, including 1,097 medium-to-high severity issues. The findings span system kernels, operating systems, browser engines, open-source infrastructure, web applications, and network protocols. Many had remained unnoticed for years or decades, with the oldest flaw dating to roughly 1981 and an average latency of 26.6 years between introduction and discovery.

A public security disclosure ledger at cvd.z.ai tracks 2,436 findings, of which 53 are publicly disclosed and 2,383 are under embargo. Severity distribution: 107 Critical, 990 High, 1,286 Medium, and 53 Low. For disclosed issues, the ledger records the affected project, severity, CVE where available, and how long the vulnerability had remained in the codebase.

Training infrastructure: slime

GLM-5.3 runs on slime, Zhipu’s open-source post-training framework for reinforcement-learning scaling, with Megatron on the training side and SGLang on the rollout side. Slime keeps training, rollout, and the data buffer on a single dataflow, so math, code, sandboxes, verifiers, and long-horizon agentic environments plug in as data generation rather than changes to the training loop. Additions through GLM-5.3 include top-p mask, top-k and full-vocabulary OPD, and configurations such as R3-style setups that improve training-rollout consistency, including full numerical alignment between training and rollout paths. Zhipu reports that the average difference in log probabilities between training and rollout was controlled at the 1e-7 level, a reduction of more than 99.99% compared with previous setups.

Availability

GLM-5.3 is available now through the GLM Coding Plan at z.ai/subscribe and works with coding agents including ZCode, Claude Code, and OpenCode. The model weights are scheduled to be released as open source two weeks after launch, once safety evaluation and hardening are complete.

FAQ

What is GLM-5.3?

GLM-5.3 is a coding-focused model from Zhipu AI released on August 14, 2026. It uses the same base model as GLM-5.2, with all reported gains coming from extended post-training.

How does GLM-5.3 compare with other coding models on benchmarks?

On Terminal Bench 3.0, GLM-5.3 scores 28.3, ahead of Claude Opus 4.8 at 21.1 and behind Claude Fable 5 at 33.7 and GPT-5.6 Sol at 34.6. On DeepSWE v1.1, GLM-5.3 reaches 66.9, behind Fable 5 at 69.7 and GPT-5.6 Sol at 72.7. Zhipu reports open-source state of the art on Terminal Bench 3.0 and Agents’ Last Exam.

When will GLM-5.3 weights be released?

Zhipu says the weights will go open source two weeks after launch, once safety evaluation and hardening are complete.

This article summarizes reporting from the-decoder.com, the-decoder.com, z.ai. See our editorial disclaimer for how our articles are produced.