GLM-5.3

BetaNews (Aug 22, 2026) reports that Z.ai announced GLM-5.3 on August 14, 2026, describing it as an open-weight model built for advanced coding and cybersecurity work, with all capability gains attributed to expanded post-training on the same base architecture as GLM-5.2. Alongside the release, Z.ai launched OpenVuln ( Concrete figures reported: CyberGym 84.5% (up from 77.2% for GLM-5.2) and ExploitBench 54.4% (more than double GLM-5.2's 24.4%); Z.ai's disclosure ledger credits GLM-5.3 with 2,436 vulnerability findings across 269 open-source projects, including 1,097 rated critical or high severity, with reports of findings in the Li

Morph's third-party profile describes GLM-5.3 as Z.ai's August 14, 2026 upgrade of the 753B GLM-5.2 mixture-of-experts base, achieved entirely through scaled post-training rather than new pretraining. The profile cites benchmark jumps of Terminal-Bench 3.0 from 4.6 to 28.3, DeepSWE v1.1 from 46.2 to 66.9, and CyberGym API pricing is reported as unchanged from GLM-5.2 at $1.40 per million input tokens and $4.40 per million output tokens, with a 1M-token context window, 128K max output, and always-on reasoning at low/high/max effort (max default). The piece also notes a public ModelOpt NVFP4 checkpoint with FP8 KV cache, deployed at t

Qubrid AI published a launch-day technical brief describing GLM-5.3—released August 14, 2026 by Z.ai—as a post-training upgrade on the identical 743B-parameter base shipped with GLM-5.2 in June, with Z.ai saying gains came from scaled post-training across more executable environments, more environment types, and longer Qubrid discloses itself as a GLM-5.3 launch partner working directly with Z.ai on the rollout, noting the API is not yet live on Qubrid pending Z.ai's safety-evaluation window but will be switched on when partner access opens. The brief also reports Z.ai's plan for open weights roughly two weeks after launch once safet

Nathan Lambert's Interconnects analysis (Aug 14, 2026) frames GLM-5.3 as Z.ai scaling post-training on the same base as GLM-5.2, at approximately 750B parameters — roughly a third of Moonshot AI's Kimi K3 — while reaching frontier-level scores on multiple agentic coding benchmarks. The article notes GLM-5.3 on several Lambert argues the simplest explanation for GLM-5.3's performance is that Z.ai is strong at post-training rather than that results are artifacts of distillation from larger US models, contrasting Z.ai's post-training strength with Kimi's pretraining-led approach. The piece also situates GLM-5.3 within broader questions

Z.ai (Zhipu AI) announced GLM-5.3 on August 14, 2026 as a post-training-only release built on the same base model as GLM-5.2, meaning every reported capability gain comes from scaled post-training rather than a re-pretraining or architecture change, according to the InferenceX technical page citing Z.ai's GLM-5.3 blog GLM-5.3 is text-only, always-reasoning, with a 1M-token context window, 128K maximum output, and three thinking-effort levels—low, high, and max (default)—with the option to disable thinking no longer supported, per the InferenceX summary of Z.ai's documentation. The page contrasts this with GLM-5.2's earlier MIT-licen