English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

GLM-5.3: Zhipu Trains Same 743B Base to Near-Frontier Coding, Uncovering 2,404 Decades-Old Vulnerabilities

Forum topic · 小凯 · 2026-08-15

Summary

On August 14, Zhipu (Z.ai) released GLM-5.3, built on exactly the same ~743B-parameter base model as GLM-5.2 — no architecture changes or added parameters — with all gains coming from post-training scaling using a new open-sourced Slime RL framework. Coding capability improved 50% over the previous generation: Terminal Bench 3.0 jumped from 4.6 to 28.3 (open-source #1, a 6x gain), DeepSWE v1.1 rose from 46.2 to 66.9 (open-source #1), and CyberGym hit 84.5%. On Z.ai Code Bench, GLM-5.3 High achieved 31.4% accuracy using only 50K tokens per task, beating Claude Opus 4.8 Max (29.5% at 120K tokens) with 2.4x token efficiency. Post-training also surfaced emergent cybersecurity skills: two weeks of pre-release testing found 2,404 potential vulnerabilities across 269 repositories (1,088 medium/high severity), some latent for over 40 years, disclosed via Zhipu's public ledger at cvd.z.ai. Full model weights will be open-sourced after a two-week hardening window.

On August 14, Zhipu released GLM-5.3. The base model is exactly the same as GLM-5.2's — still around 743B parameters, with no architecture changes and no added parameters. All improvements come from post-training scaling. Coding capability improved 50% over the previous generation: Terminal Bench 3.0 jumped from 4.6 to 28.3 (open-source #1, a 6x improvement), and DeepSWE v1.1 rose from 46.2 to 66.9 (open-source #1). On Z.ai Code Bench, GLM-5.3 at the High setting reached 31.4% accuracy, surpassing Claude Opus 4.8's Max setting at 29.5% — while consuming only 50K tokens per task, versus Opus 4.8's 120K.

Hard Evidence of Gains on an Identical Base

Zhipu was transparent about the accounting: GLM-5.3 uses the same 743B base as GLM-5.2. The post-training stage adopted a new generation of the Slime framework (the framework itself is open-sourced), introduced RL scheduling methods like IndexShare and SAO, and trained the model for an extended period in long-horizon task environments.

| Benchmark | GLM-5.2 | GLM-5.3 | Gain | |---|---|---|---| | Terminal Bench 3.0 | 4.6 | 28.3 | 6x | | DeepSWE v1.1 | 46.2 | 66.9 | +44.8% | | Agents' Last Exam | 23.8 | 28.5 | +19.7% | | CyberGym | 77.2 | 84.5 | +7.3 | | AutomationBench | 26.2 | 48.2 | +22 | | HLE w/Tools | 54.7 | 62.5 | +7.8 | | GDPval-AA v2 | 1508 | 1769 | +261 |

The model ranks open-source #1 on five of seven benchmarks and #2 on one. Terminal Bench 3.0 is the gold standard for whether a model can independently complete complex tasks in a real terminal environment: going from 4.6 to 28.3 means GLM-5.2 could barely use a terminal, while GLM-5.3 can handle work approaching an engineer's day.

Coding Efficiency: Same Work with Half the Tokens

Z.ai Code Bench runs on the Claude Code 2.1.207 harness, and the result is counterintuitive — GLM-5.3 at the High setting completes in 50K tokens what Opus 4.8 needs 120K tokens for (with 1.9 percentage points higher accuracy). That works out to 2.4x token efficiency.

Practical implication: in the API-billing era, tokens are money. GLM-5.3's 50% coding improvement is paid for through inference-stage token savings rather than user bills. Zhipu specifically called out "50K vs 120K" in its announcement, aiming directly at Opus 4.8's customer base.

Emergent Cybersecurity: 2,404 Vulnerabilities, the Oldest 40 Years Old

Post-training scaling also unexpectedly activated cybersecurity capability. Zhipu partnered with domestic security teams for two weeks of intensive pre-release testing. GLM-5.3 found 2,404 potential vulnerabilities across 269 repositories, of which 1,088 are medium-to-high severity; the oldest vulnerability had been latent for over 40 years (an architectural flaw in the DNS protocol itself, where a small number of malformed requests can amplify attacks 80,000-fold).

On CyberGym (white-box code review), GLM-5.3 scored 84.5%, beating Mythos 5 (83.8%) and GPT-5.6 Sol (83.6%) for first place overall. On ExploitBench it scored 54.4, more than double GLM-5.2's 24.4.

Zhipu launched a dedicated cvd.z.ai security disclosure ledger, publicly logging findings by CVE status, severity, and years latent; 2,383 issues remain in coordinated disclosure. This "active disclosure + delayed open-weights" cadence is unusually restrained among this wave of Chinese models.

Commercial Rollout: Lock for Two Weeks, Then Release Weights

GLM-5.3 is fully available now on ZCode, AutoClaw, and the GLM Coding Plan, with early access on developer platforms including Trae, Qoder, JoyCode, OpenCode, CodeBuddy, and WorkBuddy. Full model weights will be open-sourced in two weeks — a "hardening window" for security teams, and the first time a leading Chinese model has standardized a release-first → harden → open-source pipeline.

Two signals to watch over the next two weeks: 1) the number of community fine-tuned derivatives within 24 hours of the weight release; 2) whether GLM Coding Plan subscription growth matches Claude Code (whose weekly retention is roughly 70%).

Sources: Zhipu official announcement, Singularity.kiwi vulnerability disclosure analysis, GitHub AI Daily 8-15.

Tags

#glm-5#zhipu#zai#llm#coding-benchmarks#post-training#reinforcement-learning#cybersecurity

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178633489