Z.ai has released GLM-5.3, a coding-focused upgrade to its flagship model series that achieves substantial benchmark gains without a new pretraining run, signaling that post-training alone can deliver meaningful capability jumps.
The model, announced on August 14, reuses the roughly 700-billion-parameter base from GLM-5.2 and takes every improvement from scaled-up post-training. Terminal-Bench 3.0 jumped from 4.6% to 28.3%, while the DeepSWE v1.1 coding benchmark surged from 46.2% to 66.9%. Z.ai said thinking is now always on, with callers choosing low, high, or max effort levels.
Open Weights Coming Within Two Weeks
Z.ai plans to release the model weights publicly within two weeks of launch, allowing developers to download and modify the system. That timeline makes GLM-5.3 the latest in a string of Chinese frontier models heading toward open-weight distribution, following Alibaba’s Qwen 3.8 Max and Moonshot AI’s Kimi K3.
The company also highlighted GLM-5.3’s security capabilities. The model identified 2,436 vulnerabilities across 269 open-source projects, work Z.ai framed as defensively useful but which also raises questions about offensive capability in publicly available models.
Stock Falls Despite Model Hype
Z.ai shares closed 3.6% lower at HK$1,270 on the Hong Kong exchange on August 14, even as GLM-5.3 became a trending search topic in the United States. The stock is now 57.4% below its June peak of HK$2,980, having erased roughly HK$479 billion in market value from its all-time high. Friday’s trading range spanned 19.7%, reflecting continued volatility.
The divergence between product attention and share price underscores a shift in investor sentiment. Earlier GLM releases drove speculative rallies, but GLM-5.3 must now demonstrate commercial traction. Z.ai reported annual recurring revenue of $1 billion in July, while JPMorgan has projected revenue growth exceeding 534% in 2026 with profitability expected in 2028.
Post-Training Efficiency Matters
The GLM-5.3 release carries broader implications for the AI industry’s cost structure. A post-training-only upgrade costs far less than a full pretraining cycle, suggesting that labs can continue extracting real capability gains without the enormous compute bills associated with training new base models from scratch.
Z.ai’s technical lead Qinkai Zheng noted the tension between scaling and cost, saying the company is “trying to lower the cost, but because the demand is too large.” The company benchmarked GLM-5.3 against Anthropic’s Claude Fable 5 in several tests, though those comparisons remain company-reported and await independent verification.
Founded as Zhipu AI and listed on the Hong Kong Stock Exchange earlier this year, Z.ai has moved on a roughly two-month release cadence through the GLM-5 series: GLM-5 in February, GLM-5.1 in April, GLM-5.2 in June, and now GLM-5.3 in August.
Sources: Z.ai blog; ts2.tech; AI/TLDR daily digest; Bloomberg via The Economic Times; Reuters
discussion