Zhipu AI has released GLM-5.3, an open-weight model it calls the most capable coding model in its weight class. The release, dated August 14, 2026, uses the same base architecture as GLM-5.2, and every reported gain comes from extended post-training rather than a new pretraining run. The largest reported jumps are in agent-based coding tasks, where the model also shows an emergent cybersecurity capability that surprised the team during scaling.
What changed from GLM-5.2 to GLM-5.3
GLM-5.3 shares its base weights with GLM-5.2. According to Zhipu, all improvements come from scaling post-training on its existing stack: IndexShare for long-context processing, SAO for reinforcement learning on long-horizon tasks, and slime for asynchronous training. Over the month between releases, the team increased the number and diversity of task environments and the compute spent training on them.
The model is positioned as the strongest open-weight coding model available, with a 50% improvement over GLM-5.2 on Zhipu’s in-house Z.ai Code Bench and state-of-the-art open-weight results on Terminal Bench 3.0 and Agents’ Last Exam.
Coding benchmark results
On Terminal Bench 3.0, GLM-5.3 scores 28.3, up from 4.6 for GLM-5.2. The closed-source leaders on that benchmark, Claude Fable 5 at 33.7 and GPT-5.6 Sol at 34.6, remain ahead. On DeepSWE v1.1, GLM-5.3 reaches 66.9, compared with 46.2 for GLM-5.2 and 72.7 for GPT-5.6 Sol. On Agents’ Last Exam ALE-CLI, GLM-5.3 scores 28.5 versus 23.8 for GLM-5.2, narrowly behind GPT-5.6 Sol at 28.6 and ahead of Claude Fable 5 at 23.8.
On Z.ai’s private Z.ai Code Bench, designed to evaluate coding agents under realistic user scenarios with diverse task categories in complex local development environments, GLM-5.3 shows a 50% improvement over its predecessor. As a private benchmark, Zhipu says it reduces contamination risk from public test sets.
Task environments built to look like real engineering work
Zhipu pushed its environment scaling toward tasks that resemble real units of expert work rather than coding exercises. Some environments represent several days of work for an experienced engineer. In one ML infrastructure scenario, the model receives the same working environment as an engineer, including access to compute clusters, storage systems, internal documentation, codebases, and experiment results, and must diagnose bottlenecks across the training stack, implement optimizations, run experiments, and deliver a measurable end-to-end speedup while preserving correctness.
Research agents collect task patterns from real work and convert them into runnable long-horizon environments with multi-step dependencies and hidden state. A judge agent then attempts each task to confirm it is actually solvable. Verifiers are synthesized without access to the reference solution, and solver trajectories are used to discover and close reward shortcuts.
Token efficiency at matched effort levels
At Max effort, GLM-5.3 reaches 34.5% on Z.ai Code Bench at roughly 75,000 output tokens per task, versus 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.
An unexpected cybersecurity capability
When Zhipu introduced vulnerability discovery data and environments into the training mix, the model began to reason across multiple stages of exploitation and form coherent plans for complete exploitation chains. The capability developed faster than the team expected as post-training scaled.
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 77.2% for GLM-5.2. That places it ahead of Claude Fable 5 at 83.8% and GPT-5.6 Sol at 83.6% on the benchmark.
On ExploitBench, which requires deeper reasoning about real vulnerabilities and their exploitation, GLM-5.3 reaches 54.4%, more than doubling GLM-5.2’s 24.4%. Claude Fable 5 and GPT-5.6 Sol score 78.0% and 76.5% respectively. On ExploitGym, which measures how many exploitation tasks a model can complete 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. Claude Fable 5 remains well ahead at 181 and 247 tasks.
The pattern Zhipu highlights is consistent: 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 dating back roughly 40 years.
A public registry at cvd.z.ai tracks the findings. As of release, the registry lists 2,436 findings tracked, 53 publicly disclosed, 2,383 under embargo, 1,097 critical and high severity, across 269 open-source projects, spanning 45 years of impact. The severity breakdown shows 107 critical, 990 high, 1,286 medium, and 53 low. The oldest flaw was introduced in 1981, and on average a vulnerability lived 26.6 years before discovery. For disclosed issues, the ledger records the affected project, severity, CVE where available, and how long the vulnerability had remained in the codebase.
The slime post-training framework
All of this runs on slime, Zhipu’s open-source post-training framework for RL scaling, with Megatron on the training side and SGLang on the rollout side. The framework 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 as changes to the training loop.
Additions through GLM-5.3 include top-p mask, top-k, and full-vocabulary OPD, plus configurations improving training-rollout consistency including R3-style setups and full numerical alignment between training and rollout paths. In the training-rollout consistency evaluation, the average difference in log probabilities 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 set to go 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 released by Zhipu AI on August 14, 2026. It shares its base with GLM-5.2, and all reported gains come from extended post-training.
How does GLM-5.3 perform on coding benchmarks?
On Terminal Bench 3.0, GLM-5.3 scores 28.3, up from 4.6 for GLM-5.2. On DeepSWE v1.1 it reaches 66.9 versus 46.2. On Z.ai Code Bench it shows a 50% improvement over GLM-5.2. Closed-source leaders remain ahead on several benchmarks.
When will GLM-5.3 weights be released open source?
Zhipu plans to release the weights two weeks after launch, once safety evaluation and hardening are complete.
Related coverage
This article summarizes reporting from the-decoder.com, the-decoder.com, z.ai.