GLM-5.3

Interconnects' "Latest open artifacts" newsletter (Sep 8, 2026) covers Zhipu's GLM-5.3 as a notable licensing shift in the open-model ecosystem: GLM-5.3 moved away from the MIT license used by GLM-5.2 and earlier, adopting a custom license that imposes a condition on inference and fine-tuning providers. The clause stat The newsletter frames this change as part of a broader 2026 pattern in which Western open-model makers (Google, Meta) move toward more permissive licenses (Apache 2.0), while frontier Chinese providers like Moonshot's Kimi K3 and GLM-5.3 introduce commercial conditions, before pivoting to broader license analysis. The

BetaNews 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 that shares its base architecture with predecessor GLM 5.2, with all capability gains derived from expanded post-training. Alongside the release, Z.ai launched OpenVuln On Z.ai-reported benchmarks, BetaNews states GLM 5.3 reached 84.5% on CyberGym (up from 77.2% for GLM 5.2) and 54.4% on ExploitBench (more than double GLM 5.2's 24.4%), noting these figures are self-reported and not independently verified. Z.ai said it would hold back GLM 5.3's public model weights for about two weeks

Apidog's third-party blog (Aug 16, 2026) reports that Zhipu AI released GLM-5.3 on August 14, 2026, and frames the piece as a self-hosting preparation guide around an expected open-weights drop on Zhipu's Hugging Face organization roughly two weeks after launch — explicitly flagged by the author as an expectation rathe Apidog recommends baseline-building against the hosted API before weights go public, then replaying the same collection against a local endpoint, and notes that prior GLM-5 releases shipped on Hugging Face as paired BF16 and FP8 repos, so GLM-5.3 and a GLM-5.3-FP8 variant are expected day one, with community GGUF quant

GLM-5.3, announced on August 14, 2026 by Z.ai (Zhipu AI), is a post-training-only release that uses the same base model as GLM-5.2, with all reported gains attributed to scaled post-training. Per Z.ai's own documentation surfaced through this third-party technical write-up, GLM-5.3 ships as text-only and always-reasoni The broader GLM-5 family architecture, as summarized here from Z.ai docs, is a Mixture-of-Experts design with 744B total parameters and roughly 40B active per pass, delivered on a 1M-token context for GLM-5.2 and GLM-5.3. GLM-5.2 (Hugging Face repository created 2026-06-16) shipped under MIT with same-as-predecessor AP