GLM-5.2

A third-party technical analysis of GLM-5.2, released as a next-generation open-source coding model, details its architecture as maintaining the GLM-5 series' ultra-large parameter design with 744 billion total parameters while controlling activated parameters at approximately 40 billion through dynamic parameter activ A core architectural innovation is the DeepSeek Sparse Attention (DSA) mechanism, which addresses the O(n²) complexity of global attention through an importance evaluation module that scores tokens by syntax role, contextual relevance, and task correlation; dynamic attention allocation that preserves full computation f

NIST's Center for AI Standards and Innovation (CAISI) published an official assessment of Z.ai's GLM-5.2 on July 8, 2026, evaluating the open-weight model that was released on June 16, 2026. CAISI found GLM-5.2 was probably the most capable open-weight AI model at release, with overall capabilities similar to GPT-5.2 ( On safeguards and security, CAISI found mixed results: GLM-5.2's safeguards allow assistance with agentic cyber exploit development, and block fewer sensitive biological questions than reference U.S. models, but the model appears potentially more robust against agent hijacking and jailbreaking attacks than other evalua

Security vendor Semgrep published an independent benchmark study on June 22, 2026, in which GLM-5.2 was tested against their IDOR benchmark, the same dataset and prompt used to evaluate frontier coding agents. Among those given nothing but a prompt, the open-weight GLM-5.2 beat Claude Opus 4.8 on Semgrep's IDOR benchma The Semgrep team, including Katie Paxton-Fear, Seth Jaksik, Brenden Noblitt, and Erik Buchanan, described GLM-5.2 as the "best open-weight option" they evaluated on IDOR when compared head-to-head against Claude Opus 4.8 under matched conditions. The result is narrow in scope (a single IDOR benchmark) and carries some

Zhipu (Z.ai) released GLM-5.2 in June 2026 as the latest version of its General Language Model family, building on GLM-5 and GLM-5.1 with a clear emphasis on software engineering, long-context reasoning, and autonomous AI agents. The standout feature is support for a 1 million token context window, designed to handle r GLM-5.2 is positioned as a direct competitor to frontier coding-focused models from OpenAI, Anthropic, and Google, with Zhipu continuing to iterate rapidly on its model family. The article frames GLM-5.2 as "doubling down" on agentic software engineering, where AI systems can independently understand, plan, and execute

Semgrep evaluated GLM-5.2 against their IDOR (Insecure Direct Object Reference) benchmark using the same prompt they apply to frontier coding agents, and the result surprised the research team. Among models given nothing but a prompt, GLM-5.2 — an open-weight option — beat Claude Opus 4.8 on this cyber-security benchma The benchmark is significant because it uses identical methodology (same dataset, same prompt) previously used to evaluate frontier coding agents, making it a direct head-to-head comparison rather than a custom-tuned evaluation. This positions GLM-5.2 as a noteworthy open-weight alternative for security-focused code an