Z.ai’s new GLM-5.3-Flash model delivers near-frontier benchmark scores at roughly one-seventh the cost of its larger sibling, while running entirely on Chinese AI chips rather than Nvidia hardware. The release gives developers a 320 billion parameter, open-weight model with a one million token context window, priced at $0.09 per task on Artificial Analysis benchmarks.<\/p>\n
GLM-5.3-Flash is the first natively multimodal model in the GLM-5 series, with 320 billion total parameters, of which only 18 billion are active per pass. It ships under an MIT license and supports one million tokens of context. Weights are available on Hugging Face.<\/p>\n
On Z.ai’s API, input costs $0.15 per million tokens and output costs $0.50 per million tokens. That is a little over ten percent of the price of the larger GLM-5.3.<\/p>\n
Measurements from Artificial Analysis put GLM-5.3-Flash at 57 points on the Intelligence Index at maximum reasoning effort. That is three points behind the larger GLM-5.3, which scores 60, and level with GPT-5.6 Terra and Muse Spark 1.2.<\/p>\n
Cost per task on the index runs $0.09, against $0.68 for GLM-5.3, roughly 7.5 times cheaper. That places the model on the Pareto frontier of intelligence and cost, where the index chart shows it sitting in the most attractive cost-versus-intelligence quadrant.<\/p>\n
On agentic work the model keeps pace with bigger names. On GDPval-AA v2 it reaches an Elo score of about 1770, matching GLM-5.3 and Grok 4.6, and trailing only Claude Opus 5.<\/p>\n
There is a trade-off. Artificial Analysis found that roughly 90 percent of the output tokens GLM-5.3-Flash burned went to reasoning, which makes it less token-efficient than peers.<\/p>\n
Before launch, Z.ai tested the model anonymously as “ox-alpha” on OpenCode and OpenRouter, where it became the most popular model of the week. Every request ran on Chinese AI chips, according to Z.ai.<\/p>\n
SemiAnalysis reports the deployment served 100 trillion tokens a day, a level of capacity that until now was thought possible only for frontier labs. Z.ai puts its hardware efficiency and cost per token on par with common Nvidia GPUs.<\/p>\n
SemiAnalysis reads this as another test of the CUDA moat, the Nvidia programming layer between AI software and the graphics card that has grown for nearly 20 years and underpins nearly every tuned AI framework. Switching chips means redoing that work: reprogramming compute operations, adjusting memory access, and hunting down bottlenecks.<\/p>\n
Z.ai built its own serving software on top of SGLang and broke processing into stages that scale independently. The team says this tripled throughput over a first attempt on the same hardware. An agent based on GLM-5.3 helped with the tuning.<\/p>\n
GLM-5.3-Flash is a model released by Z.ai with 320 billion total parameters (18 billion active), a one million token context window, native multimodality, and an MIT license. Weights are available on Hugging Face.<\/p>\n
On Z.ai’s API, GLM-5.3-Flash costs $0.15 per million input tokens and $0.50 per million output tokens. On Artificial Analysis’s Intelligence Index, cost per task runs $0.09, about 7.5 times cheaper than the larger GLM-5.3.<\/p>\n
Yes. Z.ai tested the model on OpenCode and OpenRouter and ran all traffic on Chinese AI chips, with hardware efficiency and cost per token reported on par with common Nvidia GPUs.<\/p>\n