GLM open-weight model family (GLM-5 / 5.1 / 5.2 / 5.3)

Overview

Zhipu's frontier line, designed for Agentic, Reasoning, Coding workloads. GLM-5.3 (August 14, 2026) is an extended post-training pass on the same ~743B-parameter mixture-of-experts base as GLM-5.2; Hugging Face lists the published weights at 753B total parameters under a custom GLM-5.3 License. GLM-5.3-Flash (August 26) is a 320B/18B natively multimodal MIT-licensed sibling served on domestic-chip clusters.

GLM-5.3 is live on the GLM Coding Plan and ZCode, and its full open weights were published on Hugging Face on August 28, 2026 after Zhipu's internal safety review - on the promised two-week schedule, but under a bespoke GLM-5.3 License rather than GLM-5.2's plain MIT terms, requiring companies above $10B in trailing-12-month revenue to pass a Z.ai security review before commercial hosting. GLM-5.3-Flash remains MIT-licensed.

Key Specs

GLM-5.3 Base

~743B-parameter MoE base (shared with GLM-5.2), extended post-training

Coding Improvement

+50% vs GLM-5.2 on Zhipu's internal coding benchmarks

Context Window

1M tokens

Cybersecurity Benchmarks

84.5% on CyberGym (vulnerability identification from source code), narrowly ahead of Anthropic's Claude Mythos 5 (83.8%) and OpenAI's GPT-5.6 Sol (83.6%); 54.4% on ExploitBench (exploitation depth), trailing those rivals' high-70s scores. Company-reported, not independently verified; in security-team testing against real codebases GLM-5.3 flagged 2,436 vulnerabilities across 269 open-source projects.