Z.ai Official Technical Blog: Frontier Coding and Emergent Cybersecurity Capabilities (Z.ai)

Core content summary

Z.ai's official positioning: "Scaling post-training is all we did for GLM-5.3." The base model is exactly the same as GLM-5.2; all improvements come from post-training. Behind this are the IndexShare (long context), SAO (long-horizon task RL), and slime (large-scale asynchronous training) technology stack built during the GLM-5.2 period, which has continued to expand over the past month with more environments, more diverse tasks, and more training compute.

Three key release points

1. Stronger coding capabilities: The official announcement calls GLM-5.3 "the strongest open-weight coding model"; its internal Z.ai Code Bench score is 50% higher than 5.2, while Terminal Bench 3.0 and Agents' Last Exam reach open-source SOTA.

2. Emergent cybersecurity capabilities: As the scale of post-training grew, cybersecurity capabilities improved beyond expectations. Vulnerability discovery on CyberGym reached SOTA, with more pronounced gains the deeper the model went along the exploit chain (scores on vulnerability-exploitation benchmarks more than doubled).

3. Open-sourcing: The weights were released about two weeks later, after safety evaluation and hardening were completed.

Official full benchmark table (vendor-reported)

Key training methodology

• Scaling environments: Training environments shifted from "coding problems" to "complete professional work units"—some tasks are equivalent to several days of work by a senior engineer (for example, an ML infrastructure task involving access to compute clusters, storage, internal documentation, code repositories, and experiment results; diagnosing bottlenecks, implementing optimizations, running experiments, and delivering measurable end-to-end speedups).

• Automated environment-synthesis pipeline: A research agent collects task patterns from real work and generates runnable long-horizon environments; a review agent verifies that tasks are solvable; and a verifier generates solutions without accessing the reference answer, checking oracle / no-op / unresolved states to produce reliable binary rewards that can be used directly for training.

• SAO with compaction: The RL strategy from GLM-5.2 continues, ensuring that gains persist on long-horizon tasks.

• Efficiency data: The Max tier achieves 34.5% @ ~75K tokens versus 23.4% @ ~96K for 5.2; the High tier reaches 31.4% @ ~50K, exceeding Opus 4.8's 29.5% @ ~120K; it still trails Fable 5 (39.5% in the Max tier).

Cybersecurity and disclosure ledger

• Data: 2,436 vulnerability discoveries / 269 open-source projects / 1,097 critical or high-severity vulnerabilities; the earliest vulnerability was introduced in 1981, with an average "dwell time" of 26.6 years; severity distribution: Critical 107, High 990, Medium 1,286, Low 53.

• Z.ai security disclosure ledger: https://cvd.z.ai/

• "capability is growing fastest exactly where we are furthest behind"—capabilities are growing fastest precisely where the gap is largest.

slime framework and engineering details

• The open-source post-training framework slime (Megatron on the training side, SGLang on the rollout side) uses a single data flow for training, rollout, and data buffering; additions include top-p masking, top-k and full-vocabulary OPD, an R3-style configuration, and training-rollout numerical alignment (the average logprob difference is controlled at around 1e-7, more than 99.99% lower than before); a local-storage caching layer, dynamic teacher-switching prefetch, and joint routing-and-slime scheduling raise end-to-end throughput for long-horizon coding RL by more than 2.3x.

API changes (official)

• Reasoning is now mandatory: thinking.type supports only enabled; reasoning_effort has three tiers—low/high/max (max by default; max is recommended for coding).

• Migration: Applications that previously used thinking.type:"disabled" must switch to enabled + reasoning_effort:"low" before changing the model ID, or the request will fail.

• Example request body: {"model":"glm-5.3","thinking":{"type":"enabled"},"reasoning_effort":"max"}

• GLM Coding Plan has switched to a credit-quota system; during peak hours (Monday through Friday, 14:00–18:00 UTC+8), external calls consume 50% of the standard credit quota.

• ZCode: 98%+ cache hit rate (about 30% more effective tokens), a 1.5x limited-time quota bonus (through August 31, up to 180% when stacked), Goal mode, and remote control through WeChat / Feishu.