What Zhipu AI is up to now

New 08-31 PM: the afternoon re-check holds the GLM-5.5 negative and adds a finding about the expectation's own framing. The zai-org Hugging Face org carries GLM-5.3 and its Flash variants with checkpoint uploads hours old, and no GLM-5.5 model card exists anywhere checked; Chinese coverage says the numbering stayed on the 5.x track and that "5.5" was a rumoured designation Zhipu never used. This does NOT mean GLM-5.3 satisfies the expectation: the claim is a >1T-parameter, 1M-context model and GLM-5.3 is 753B. So the claimed model does not exist under any name, and the label it was logged under appears abandoned by the vendor. New 08-31: GLM-5.5 still has not shipped, and today's negative is stronger than yesterday's. Neither Hugging Face nor Z.AI's own documentation carries it — while the company is demonstrably publishing this morning, shipping GLM-5.3-Flash updates. So this is a live vendor channel shipping something else, not an unreachable one; yesterday's check rested on secondary sources after a DNS failure on z.ai. The `glm-5-5-release` expectation is due today and is held pending until the window closes. New 08-15→08-16: still quiet — both days' only Zhipu/GLM mentions are explicit cross-references back to the 08-14 GLM-5.3 launch already captured below (08-16's TLDR AI Monday-lead item is confirmed weekday-publication lag on the same 08-14 story, not a new event). No movement yet on the ~08-28 Hugging Face weight release or the pending, later GLM-5.5 (due 08-31). Standing picture below holds. New 08-14 (the only substantive development found across the full ~3-week gap since 07-24): Zhipu, operating as Z.ai, launched its flagship GLM-5.3, claiming it edges out Anthropic's restricted-access Claude Mythos 5 on the CyberGym cybersecurity benchmark (84.5% vs. 83.8%) while trailing badly on ExploitBench (54.4% vs. 78%). Runs on the same 743B-parameter MoE base as GLM-5.2 (~40B active parameters/token). Notably, Zhipu is holding the weights back from Hugging Face until roughly 08-28 for its own safety review, with the most sensitive cybersecurity functions gated behind a "trusted access" program for verified users — the first time a Chinese open-weight lab has voluntarily gated a release this way, mirroring the access controls Anthropic itself uses for Mythos. Distinct from the still-pending, later GLM-5.5 release (also due 08-31) — not a flip of that entry. Standing picture unchanged: the 1GW all-domestic-chip training site (Z.AI) and the GLM line remain the clearest proof China can train at frontier scale without Nvidia; GLM-5.2 was the model Hugging Face used to investigate OpenAI's containment breach. New 08-26: confirmed it built "Ox Alpha," the stealth reasoning/coding model released uncredited over the previous weekend that had climbed past DeepSeek to top OpenRouter usage, and said it would open the weights the same night. Anonymity was the launch tactic: the model earned its leaderboard position with no national label attached, and the label arrived only once the position was unarguable.

as of 2026-08-31