GLM-5.3 open-weight release
Archived — this story has rotated out of today’s deck. It is kept here in full.
The gist
Z.ai released GLM-5.3 open-weight model on Aug 28, 2026, after a two-week delay. The release intensifies the open vs. closed AI debate and tests China's domestic chip scaling.
Background
Z.ai, a Chinese AI lab, announced GLM-5.3 on Aug 14, 2026, as an open-weight model for coding and cybersecurity. It delayed public release by two weeks due to cybersecurity risks, granting early access to vetted partners. The model is built on GLM-5.2's architecture with expanded post-training, and its release comes amid a broader debate on open-weight AI safety and China's push for domestic chip independence.
How it unfolded
- Aug 14, 2026Z.ai announced GLM-5.3, an open-weight AI model for coding and cybersecurity, with a two-week delay on public weights.
- Aug 22, 2026Z.ai confirmed the delay, citing cybersecurity risks, and restricted early access to vetted security partners.
- Aug 25, 2026The New York Times reported on the planned release, highlighting the open vs. closed AI debate.
- Aug 26, 2026Z.ai revealed it was behind the stealth model Ox Alpha, which became GLM-5.3-Flash, and released its weights.
- Aug 27, 2026Zhipu AI shares jumped 12% as GLM-5.3-Flash was revealed to run on 100,000 domestic chips, processing 62 trillion tokens.
- Aug 28, 2026Z.ai released the flagship GLM-5.3 open-weight model, with io.net as a Day 0 launch partner.
Who’s saying what
- Official
- Z.ai stated the model's gains 'can help defenders identify weaknesses earlier, validate risks, and accelerate remediation.'
- Caution
- Some experts argue that open-weight AI models are too dangerous to share publicly, citing risks of misuse.
- Support
- Dan Lahav, CEO of Irregular, said 'Open weight models have a very important part to play.'
- Analysts
- Analysts note the release tests China's ability to scale AI on domestic chips, challenging Nvidia's dominance.
Still unverified
The claim that GLM-5.3 rivals Anthropic's Fable 5 on certain benchmarks is based on Z.ai's own statements and not independently verified.