Zhipu AI released GLM-5.3, delivering a 50% improvement in coding performance over GLM-5.2 through scaled post-training alone.

The Chinese AI lab behind ChatGLM used the same base model as GLM-5.2, with all gains coming from enhanced post-training on long-horizon task environments. The company spent the past month scaling its training stack across more diverse environments and increased compute.

Coding Performance Jumps

GLM-5.3 achieves state-of-the-art results among open-weight models on multiple coding benchmarks. The model scored 28.3 on Terminal Bench 3.0, up from 4.6 in GLM-5.2, and improved from 46.2 to 66.9 on DeepSWE v1.1.

On Zhipu AI's internal Z.ai Code Bench, GLM-5.3 reached 34.5% task completion at maximum effort while using 75,000 output tokens per task. GLM-5.2 achieved only 23.4% completion using 96,000 tokens.

The training focused on production-level workflows that mirror real engineering work. Tasks now span several days of expert work, including ML infrastructure optimization where models must diagnose bottlenecks, implement fixes, and deliver measurable performance improvements.

Unexpected Cyber Capabilities

The model developed cybersecurity skills faster than anticipated during training. GLM-5.3 scored 84.5 on CyberGym for vulnerability discovery, compared to GLM-5.2's 77.2.

More significantly, the model began reasoning across complete exploitation chains rather than identifying isolated vulnerabilities. On ExploitGym benchmarks, GLM-5.3 more than doubled GLM-5.2's performance, scoring 105 versus 29 on the 2-hour test and 130 versus 39 on the 6-hour evaluation.

Training Infrastructure

Zhipu AI built pipelines that synthesize training environments end-to-end. Research agents collect task patterns from real work and convert them into executable long-horizon environments with multi-step dependencies.

The company uses its SAO reinforcement learning approach with compaction to maintain performance gains on extended tasks. Judge agents verify that synthesized tasks are solvable, while verifiers generate binary rewards for direct training.

Release Timeline

Zhipu AI will open-source GLM-5.3 weights in two weeks following safety evaluation and hardening. The model is currently available through the company's Z.ai API and ZCode development environment.

The release puts GLM-5.3 ahead of most closed models on coding benchmarks, though it remains behind Anthropic's Claude Fable 5, which achieved 39.5% on the Z.ai Code Bench at maximum effort.