GLM-5.3

NIST's CAISI published an assessment of Z.ai's GLM-5.3 on September 17, 2026, confirming that Z.ai released the model on August 14, 2026, with open weights following two weeks later. CAISI's key findings are that GLM-5.3 is the most cyber-capable open-weight model released to date, yet its cyber capabilities are signif CAISI evaluated GLM-5.3 on four benchmarks covering vulnerability discovery and exploit development, reporting higher success rates as indicating greater cyber capability, with Wilson 95% confidence intervals shown in Figure 2. CAISI notes it had previously assessed Z.ai's GLM-5.2, providing a comparison point. The ass

Interconnects' "Latest open artifacts (#24)" newsletter (September 8, 2026) covers GLM-5.3's release and notes a significant licensing shift: unlike GLM-5.2 and earlier GLM models that used MIT, GLM-5.3 ships under a custom license. For Model-as-a-Service providers whose aggregate revenue (including affiliates) exceeds The piece flags that "affiliates" is undefined in the English license text, creating uncertainty and adoption barriers, while the Chinese version uses "关联方," which has a defined meaning under Chinese law. It situates GLM-5.3 alongside other 2026 licensing trends: Google and Meta adopting Apache 2.0, while frontier C

Z.ai announced GLM 5.3 on August 14, 2026 as an open-weight model built for advanced coding and cybersecurity work, running on the same base architecture as GLM 5.2 with all capability gains coming from expanded post-training rather than a larger base model. On Z.ai's CyberGym vulnerability-finding benchmark, GLM 5.3 r Z.ai's public 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 several flaws reportedly found in the Linux kernel and widely used VMware and Apache projects; a Z.ai developer advocate also posted on X that GLM 5.3

Z.ai released GLM-5.3 on August 14, 2026, as a coding- and cybersecurity-focused model that keeps the GLM-5.2 base model unchanged and derives all reported improvements from post-training, per DataNorth. It ships with a 1,000,000-token context window and a 128,000-token output limit, and on Z.ai's own evaluations lifts The model is text-in/text-out (no vision or audio path) and exposes three reasoning effort levels — low, high, and max — with max as the default, and no longer supports disabling thinking; requests that pass a thinking type of disabled will fail rather than fall back. DataNorth frames GLM-5.3 as targeted at long-horizo

According to the TokenPlan comparison page (last updated 2026-08-15), Volcengine Ark Coding Plan routes GLM-5.3 alongside Doubao Seed 2.0/2.1, GLM-5.2, Kimi-K3, Kimi-K2.7-Code, and DeepSeek-V4-Flash/V4-Pro under a single shared Coding Plan Key. The page states GLM-5.3 launched in August, though it does not distinguish The page describes two subscription tiers: Lite at ¥40/mo (1,200 calls per 5h, 9,000 per week, 18,000 per month) and Pro at ¥200/mo (6,000 per 5h, 45,000 per week, 90,000 per month), with promotional rates of approximately ¥9.9/¥49.9 for the first two months. Quota is shared across IDE clients including TRAE, Cursor, C

South China Morning Post reports that Zhipu (Z.ai) unveiled GLM-5.3 as its flagship model on August 14, 2026, positioning it within China's push for AI cyber-defence capability. The article states Z.ai claimed GLM-5.3 outperformed leading US systems on cybersecurity tests, specifically citing beats over Anthropic's Myt The SCMP piece frames GLM-5.3 against the broader US-China AI race and Western frontier model advances, noting its cybersecurity benchmark positioning. The supplied excerpt is heavily truncated by the SCMP subscription/paywall, so technical specifications (context length, parameters, benchmarks beyond the headline clai

GLM-5.3, announced by Z.ai on August 14, 2026, is a post-training-only update that reuses the GLM-5.2 base model, with Z.ai stating every gain comes from post-training. According to the Z.ai GLM-5.3 model card and docs, it is a text-only, always-reasoning model with a 1M-token context, a 128K max output, and three effo The candidate page also highlights GLM-5.3's licensing shift: unlike GLM-5.2's MIT license, GLM-5.3 uses a custom license that imposes a security-review requirement on Model-as-a-Service providers whose aggregate revenue (including affiliates) exceeds US$10 billion over any consecutive 12 months. The article notes that