Z.ai runs GLM-5.3-Flash on 100 000 Chinese AI chips instead of Nvidia GPUs

Z.ai confirmed that 100 000 Chinese-made chips, rather than Nvidia graphics cards, handle GLM-5.3-Flash operation for online queries. According to the company, its own software stack built on SGLang achieved efficiency on these chips comparable to standard Nvidia GPUs and tripled throughput compared with the first attempt on the same hardware; an agent built on GLM-5.3 also contributed to the optimization. SemiAnalysis reported that the model can serve 100 trillion tokens per day this way. The move follows efforts by China to reduce dependence on Nvidia - in September 2025, the Chinese government ordered Alibaba and ByteDance to stop buying Nvidia AI chips, and in November it banned foreign AI chips in state-funded data centers.

Z.ai initially tested GLM-5.3-Flash under the codename Ox Alpha from 20 August 2026, officially releasing it on 26 August 2026 as an open-weight model under the MIT license. It is the first natively multimodal model in the GLM-5 series, with 320 billion parameters, of which 18 billion are active, and a context window of 1 million tokens; the weights are available on Hugging Face. According to Z.ai, the model focuses on long contexts, visually guided agentic tasks and code synthesis.

According to measurements by Artificial Analysis, the model achieves 57 points in the Intelligence Index at the maximum reasoning setting, just three points below the larger GLM-5.3 (60 points) and comparable to GPT-5.6 Terra and Muse Spark 1.2. The cost per task according to the same index is 0.09 dollars compared with 0.68 dollars for GLM-5.3, roughly 7.5 times lower. Through the API from Z.ai, the model costs 0.075 dollars per million input tokens and 0.25 dollars per million output tokens until 9 September 2026, after which prices rise to 0.15 and 0.50 dollars per million tokens, respectively. For comparison, GPT-5.6 Luna from OpenAI costs 0.20 dollars per million input tokens and 2 dollars per million output tokens, while Claude Opus 5 from Anthropic costs 5 dollars per million input tokens and per million output tokens. On the GDPval-AA v2 benchmark, GLM-5.3-Flash achieves an Elo score of around 1770, equal to GLM-5.3 and Grok 4.6, and trails only Claude Opus 5. According to Artificial Analysis, approximately 90 percent of the output tokens from the model go toward reasoning, making it less token-efficient.