GLM-5.3: Post-Training Drives Major Gains in Coding and Cyber Capabilities

On August 14, Zhipu AI released GLM-5.3. The model shares the same base as GLM-5.2 — every improvement comes from post-training reinforcement learning. Coding performance improved by 50%, and the model achieved open-source SOTA on cyber vulnerability discovery benchmarks.

Same Base, Scaled Post-Training

GLM-5.3 uses the exact same foundation as GLM-5.2. Zhipu AI's approach was to scale up post-training with more task environments, more diverse tasks, and more compute. Their framework, slime, integrates Megatron for training and SGLang for inference, allowing math, code, sandboxes, verifiers, and long-horizon environments to plug in as data generation rather than changes to the training loop.

From GLM-5.2 to GLM-5.3, Zhipu AI expanded post-training environments in three directions. First, pushing tasks from coding exercises toward production-level work that mirrors what experienced engineers actually do. Second, introducing vulnerability discovery data into training — which produced unexpectedly strong cyber capabilities. Third, maintaining the SAO strategy from GLM-5.2 for long-context reinforcement learning with compaction.

Coding: New Open-Source Benchmark Leader

GLM-5.3 achieved the best open-source scores across multiple coding benchmarks:

Zhipu AI also introduced Z.ai Code Bench, an internal benchmark evaluating agents under realistic user scenarios. At Max effort, GLM-5.3 reaches 34.5% completion with roughly 75K output tokens, compared to GLM-5.2's 23.4% at 96K tokens. Claude Opus 4.8 scores 29.5% at 120K tokens, and Claude Fable 5 reaches 39.5%.

The token efficiency gains are notable: GLM-5.3 uses fewer tokens to achieve higher scores, suggesting the post-training improvements extend beyond raw capability to inference efficiency.

Cyber Capabilities: From Identification to Exploitation Chains

The most surprising improvement appeared in cybersecurity. Zhipu AI introduced vulnerability discovery data during post-training to help the model identify and reason about vulnerabilities. What emerged went further than expected — the model began reasoning across multiple exploitation stages, forming coherent plans for complete attack chains.

Benchmark results:

The pattern is consistent: the closer a benchmark gets to real attack chains, the larger GLM-5.3's gains over GLM-5.2 — and the wider the gap to closed frontier models. The capability grows fastest exactly where it's furthest behind.

In real-world testing, GLM-5.3 found 2,436 vulnerabilities across 269 open-source projects, including 1,097 medium-to-high severity issues. The oldest vulnerability dated back to 1981, and on average vulnerabilities had been hiding for 26.6 years before discovery.

Availability and Pricing

GLM-5.3 is available via Z.ai API and ZCode IDE. It's also accessible on OpenRouter at $1.40/M input tokens and $4.40/M output tokens, with a 1M token context window.

The model offers three reasoning effort levels: low, high, and max. High is recommended for coding tasks. Disabling thinking is no longer supported.

Zhipu AI plans to release model weights approximately two weeks after launch, following safety evaluation and hardening.

Takeaway

GLM-5.3 demonstrates that post-training scaling on a frozen base model still has significant room for improvement. A 50% coding leap, unexpected cyber capability emergence, and better token efficiency — all from the same technical direction of training on more, harder environments.

The emergent cyber capabilities deserve particular attention. When a model starts reasoning through complete exploitation chains, the offensive-defensive landscape may be shifting. Zhipu AI has already identified 2,436 vulnerabilities across 269 open-source projects, with 53 publicly disclosed so far. That number continues to grow.