GLM-5.2

A Baidu Cloud international article dated September 7, 2026 describes GLM-5.2's technical architecture, reporting that it maintains 744 billion total parameters with approximately 40 billion active parameters using a "sparse activation" strategy that reduces inference computational requirements by 42%. The article attr The article claims GLM-5.2 was officially released on June 13, 2026 (conflicting with the CAISI assessment's June 16 date), and highlights performance gains including a 67% reduction in attention computation while maintaining 98% inference accuracy, with only 18% increased memory usage when handling long contexts. The

SaferAI published an independent risk evaluation report on August 2, 2026 (authors Chinmayi Dixit, Jacob Davies, Jasmine Li, Ben Snodin, Jack Kengott, Henry Papadatos) assessing GLM-5.2, Z.ai's open-weight flagship released June 16th 2026, against the four systemic risk areas defined in the EU General-Purpose AI Code o On specific axes, SaferAI found GLM-5.2 roughly level with Opus 4.7 and slightly below GPT-5.5 on biological knowledge (~two months behind the frontier), around 2-4 months behind on cyber (at the level of Opus 4.6 and near GPT-5.5), and the furthest behind on software engineering (below Opus 4.6 and GPT-5.4, the fronti

NIST's Center for AI Standards and Innovation (CAISI) published an independent assessment of Z.ai's GLM-5.2 on July 8, 2026, concluding it was "probably the most capable open-weight AI model when it was released" on June 16, 2026. According to CAISI's evaluations, GLM-5.2's overall capabilities are similar to GPT-5.2 ( CAISI's safeguard evaluation 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 it appears potentially more robust against agent hijacking and jailbreaking attacks than other evaluated PRC open-

Zero Gravity Labs announced on June 18, 2026 (distributed via Yahoo Finance/GlobeNewswire as a paid press release) that Z.ai's GLM-5.2 is live on 0G Private Computer, enabling developers to run the model on fully private, verifiable infrastructure. The release restates Z.ai's published specifications: GLM-5.2 is open-w The 0G announcement also surfaces Z.ai's developer-stated benchmark claims for GLM-5.2: it is reported as the first open-weight model to exceed 80% on Terminal-Bench 2.1, to trail Anthropic's Claude Opus 4.8 by about one point on FrontierSWE, and to match or beat GPT-5.5 on several long-horizon coding tasks at a fracti

Featherless published a Day Zero launch partner post on June 18, 2026, confirming GLM-5.2 shipped on June 16, 2026 and was available on its platform the same day. GLM-5.2 is described as roughly 753B parameters in a Mixture-of-Experts design activating around 39B per token, with gains over GLM-5.1 coming from a revised On coding benchmarks, GLM-5.2 is reported as the highest-ranked open-source model across FrontierSWE, PostTrainBench, and SWE-Marathon and the only open-weight model ranking alongside Claude Opus 4.8 and GPT-5.5 on that class of work. Specific scores cited include Terminal-Bench 2.1 at 81.0, SWE-bench Pro at 62.1, Fron

Z.ai announced GLM-5.2 on June 16, 2026 as its flagship model built for long-horizon tasks, marking a substantial capability leap over predecessor GLM-5.1 and, for the first time, delivering that capability on a solid 1M-token context. The release highlights Advanced Coding with Flexible Effort (multiple thinking effor Architecturally, Z.ai introduces IndexShare, which reuses the same indexer across every four sparse attention layers, reducing per-token FLOPs by 2.9x at 1M context length, alongside an improved MTP layer for speculative decoding that increases acceptance length by up to 20%. On long-horizon coding benchmarks, GLM-5.2

BenchLM's model record for GLM-5.2, updated through September 4, 2026, marks it as superseded by the newer GLM-5.3 in Z.ai's line. The profile shows GLM-5.2 scoring 68.2 out of 100 and ranking 30 of 232 tracked models, with API pricing fixed at $1.40 per million input tokens and $4.40 per million output tokens. It repo Within BenchLM's eligible categories, GLM-5.2 places in the 90th percentile for Coding (rank 20 of 183), 84th percentile for Agentic work (rank 25 of 151), and 79th percentile for Knowledge (rank 38 of 181); Reasoning, Math, Multilingual, Multimodal, and Instruction Following are not rank-eligible. The page also flags

Semgrep published an independent benchmark blog on June 22, 2026 titled "We have Mythos at Home: GLM 5.2 beats Claude in our Cyber Benchmarks," in which researchers Katie Paxton-Fear, Seth Jaksik, Brenden Noblitt, and Erik Buchanan evaluated several open-source models against Semgrep's IDOR benchmark using the same dat The Semgrep piece is positioned within the vendor's Security Research content alongside broader Semgrep product navigation (Code, Supply Chain, Secrets, Guardian, AppSec Platform). While the article's core technical contribution is the IDOR benchmark comparison against GLM 5.2 and Claude Opus 4.8, the page also contain