GLM-5.2

SaferAI published an independent external risk evaluation of GLM-5.2 on August 2, 2026 — the first such European evaluation — testing Z.ai's open-weight flagship (released June 16, 2026) across the four systemic risk areas in the EU General-Purpose AI Code of Practice: Loss of Control, Cyber Offense, CBRN, and Harmful The report further notes that on cyber capability GLM-5.2 was around 2–4 months behind the frontier, at the level of Opus 4.6 and near GPT-5.5, and on software engineering it was the furthest behind, below Opus 4.6 and GPT-5.4. Beyond capability, SaferAI observed that GLM-5.2 refused none of the offensive-security or b

The U.S. National Institute of Standards and Technology's Center for AI Standards and Innovation (CAISI) published an independent assessment of Z.ai's GLM-5.2 on July 17, 2026, evaluating the open-weight model released June 16, 2026. CAISI found that GLM-5.2 was probably the most capable open-weight AI model at release CAISI's safeguards assessment was mixed: 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 appeared potentially more robust against agent hijacking and jailbreaking attacks than other evaluated PRC open-w

Lambda published a third-party analysis on July 9, 2026 noting that Z.ai released GLM 5.2 on June 16, 2026 with advertised scores at or near Anthropic and OpenAI flagships and a substantial leap over GLM 5.1. The piece explicitly names GLM 5.2 and confirms it is the same architecture as the prior GLM family at 744 bill Lambda frames GLM 5.2 as a potential "DeepSeek moment for agents," reporting that reputable sources and industry labs are replacing significant portions of their Claude and Codex workloads with the model after extended testing, citing anecdotal signals such as Databricks MTS Yuchen noting increased corporate demand for

Chinese AI lab Z.ai released GLM-5.2 to its coding plan subscribers on June 13, 2026, followed by a full open-weights release under an MIT license on June 16, 2026. The model is a 753B-parameter Mixture-of-Experts architecture with 40 active parameters, totaling 1.51TB of weights. It supports a 1 million token context GLM-5.2 tops the Artificial Analysis Intelligence Index v4.1 among open-weights models with a score of 51, ahead of competitors scoring 43–44, though it is notably token-hungry, using 43k output tokens per benchmark task versus 24k–37k for peers. It ranks #2 on the Code Arena WebDev leaderboard for front-end and agenti

Z.ai announced GLM-5.2 on June 16, 2026 as its latest flagship model purpose-built for long-horizon tasks, delivering what it describes as a "solid 1M-token context" that stably sustains extended agentic coding work rather than merely accepting more tokens. The release post explicitly names GLM-5.2 (not a sibling or fa For long-horizon coding benchmarks, the announcement claims GLM-5.2 trails Anthropic's Opus 4.8 by only 1% on FrontierSWE while edging out GPT-5.5 by 1% and Opus 4.7 by 11%; on PostTrainBench (where each agent is given an H100 to post-train small models) GLM-5.2 ranks second only to Opus 4.8, outperforming both Opus 4.

Z.ai's developer documentation release notes entry dated 2026-06-16 explicitly names GLM-5.2 and summarizes its capabilities: support for 1M lossless context that reduces context drift and goal forgetting in complex tasks, open-source SOTA performance on coding and long-horizon task benchmarks, and improved real-world As a first-party Z.ai developer-facing source, the release notes complement the announcement blog by providing a concise capability checklist aimed at API and integration users rather than researchers. The page confirms that GLM-5.2 was the shipped version on June 16, 2026 and frames it as the long-horizon, large-conte

Semgrep published a security-research post on June 22, 2026 titled "We have Mythos at Home: GLM 5.2 beats Claude in our Cyber Benchmarks," reporting that on their IDOR benchmark — the same dataset and prompt previously used to evaluate frontier coding agents — GLM 5.2 was the best-performing open-weight model they test Semgrep's findings position GLM 5.2 as a serious open-weight option for security-focused coding tasks where the model is given nothing but a prompt, rather than access to rule-based tooling. The article is authored by Semgrep researchers (Katie Paxton-Fear, Seth Jaksik, Brenden Noblitt, Erik Buchanan) and sits within t