GLM-5.3

Z.ai released GLM-5.3 on August 14, 2026 as a post-training-only update built on the same base model as GLM-5.2, with all reported gains derived from scaling post-training on the existing stack (IndexShare, SAO, slime). According to the Z.ai launch blog, GLM-5.3 delivers a 50% improvement over GLM-5.2 on Z.ai's in-hous The supplied benchmark table compares GLM-5.3 against GLM-5.2, Kimi K3, DeepSeek-V4 Pro-0813, Qwen3.8-Max, Opus 4.8, Fable 5 (with fallback), and GPT-5.6 Sol across coding, cyber, and agentic suites. Notable figures from the page include Terminal Bench 3.0 at 28.3 (vs. 4.6 for GLM-5.2), DeepSWE v1.1 at 66.9, FrontierSW

On September 17, 2026, NIST's Center for AI Standards and Innovation (CAISI) published an independent assessment of GLM-5.3's cyber capabilities, concluding that GLM-5.3 is the most cyber-capable open-weight model released to date, but still significantly below current U.S. frontier models. CAISI's aggregate measure ac CAISI evaluated GLM-5.3 on four benchmarks covering vulnerability discovery and exploit development, building on its prior assessment of Z.ai's GLM-5.2. The assessment provides an authoritative external counterpoint to Z.ai's launch claims of emergent cyber capability and SOTA CyberGym performance, and is directly rele

BetaNews reports that Z.ai announced GLM-5.3 on August 14, 2026 as an open-weight model built for advanced coding and cybersecurity, releasing OpenVuln (branded "VulnHunter"), a scanning service that uses GLM-5.3 to check public code repositories for security flaws while keeping detailed findings private until fixes ar Z.ai said it would hold back GLM-5.3's public model weights for about two weeks after launch, restricting access during that window to vetted security partners through its GLM Coding Plan and ZCode agent, while acknowledging in its announcement that the same capability gains that help defenders also carry risk for atta

Morph's August 21, 2026 analysis characterizes GLM-5.3 as Z.ai's August 14, 2026 upgrade of GLM-5.2 under the tagline "Built to Code. Ready for Cyber Defense.": the same 753B MoE base, no new pretraining, with all gains from scaled post-training focused on agentic coding, long-horizon tool use, and vulnerability resear On benchmarks, Morph reports vendor figures Terminal-Bench 3.0 4.6 to 28.3, DeepSWE v1.1 46.2 to 66.9, and CyberGym 77.2% to 84.5%, and cites an independent Artificial Analysis Intelligence Index score of 60 (rank 8 of 182, tying Kimi K3 at roughly one-fifth the price). API pricing is unchanged at $1.40 per 1M input an

MarkTechPost's August 14, 2026 recap reports that Z.ai released GLM-5.3 running on the same 743B base model as GLM-5.2, with all reported gains coming from scaled post-training involving more task environments, more environment types, and longer training. The coding improvements land most on the longest-horizon benchma The deployment analysis describes GLM-5.3 as partially deployable: live through the Z.ai API, the GLM Coding Plan, and ZCode, but with no weights available yet. Startups and mid-market engineering organizations can adopt it immediately via the Coding Plan or API, while enterprises with data-residency or vendor-review r

Nathan Lambert's Interconnects analysis, published August 14, 2026, examines Z.ai's GLM-5.3 and characterizes it as an exceptional release, noting that on many benchmarks the model surpassed Moonshot AI's Kimi K3 and on some benchmarks exceeded Claude Fable 5 or GPT-5.6-Sol. Lambert highlights that this places GLM-5.3 Lambert notes that GLM-5.3 was initially available only in the Z.ai coding plan at launch, with API access coming soon and open weights to follow in approximately two weeks on Hugging Face. He adds an analytical bullet on the Chinese data industry context for these results, arguing that the simplest explanation for how

The InferenceX/SemiAnalysis deep dive confirms GLM-5.3 is a post-training-only release announced on August 14, 2026, using the same base model as GLM-5.2. The page documents GLM-5.3's specifications as text-only, always-reasoning, with a 1M-token context, 128K max output, and three thinking-effort levels (low, high, ma The analysis reports that working with security teams in China, Z.ai found GLM-5.3 surfaced 2,436 vulnerabilities across 269 projects, including 1,097 medium-to-high severity issues, with an average latent lifetime of 26.6 years and the oldest flaw introduced in 1981. The page distinguishes licensing and availability b

Z.ai shipped GLM-5.3 on August 14, 2026 as the latest open-weight model in the GLM series, positioned as a frontier coding and agentic model. According to the Eigent writeup, the release is unusual for a point release: the base model is unchanged from GLM-5.2 — a 743B-parameter Mixture-of-Experts base per this source ( On vendor-reported coding benchmarks, Eigent lists Terminal-Bench 3.0 jumping from 4.6 to 28.3 (roughly 6x on the longest-horizon CLI benchmark), DeepSWE v1.1 from 46.2 to 66.9, Agents' Last Exam (CLI) from 23.8 to 28.5, and GDPval-AA v2 scoring 1,769 across 44 occupations; Z.ai also reports a 50% figure on its interna