Zhipu Releases 743B-Parameter GLM-5.3 with Major Leaps in Coding and Security
智东西 · wechat · 2026-08-14
Zhipu has officially released its latest flagship model, GLM-5.3, boasting 743 billion parameters with plans to open-source it within two weeks. The model's coding and agentic capabilities closely approach top-tier models like Claude Opus 4.8. Under a similar thinking budget, GLM-5.3 outperforms Opus 4.8 in accuracy while consuming less than half the tokens.
In hands-on tests, GLM-5.3 paired with Zhipu's ZCode tool successfully developed a 3D open-world driving game using real map data. It also parsed the 7,000+ file DeepSeek Harness codebase from scratch to build a 'persona plugin.' Furthermore, driven by post-training scaling, GLM-5.3 demonstrated emergent cybersecurity capabilities, uncovering 2,436 vulnerabilities during a two-week red-teaming exercise.
More from coding & agent
- AI Agent Specula Finds 249 Bugs, but Expert Questions Its Approach — tianyin_xu · 2026-08-14
- Open-Source MCP Server Bridges Claude Code and OpenAI Codex CLI — tom_doerr · 2026-08-14
- Replace $2K/mo n8n Stack with a $5 Cloudflare Worker Using AI Coding Tools — thisdudelikesAI · 2026-08-14
- System Prompts Are Not Access Control: Building Secure AI Employee Architectures — Mahmoud_Zalt · 2026-08-14
- GLM-5.3 Fully Tested: Massive Leaps via Post-Training Scaling, Best Open Model? — WorldofAI · 2026-08-14
- Best AI Coding Subscription Under $20: Adjusting Cost for Subsidies & Promos — popiazaza · 2026-08-14