Z.ai has launched GLM-5.3-Flash, a 320 billion parameter model capable of handling one million tokens in a single context window.
The release is notable for two reasons. It approaches the performance of its larger sibling while costing significantly less, and it runs entirely on Chinese AI hardware rather than Nvidia GPUs.
Performance and pricing
According to Z.ai, this is the first natively multimodal model in the GLM-5 series. It uses an MIT license and the weights are available on Hugging Face.
Testing by Artificial Analysis shows the model scores 57 points on the Intelligence Index at maximum reasoning effort. This places it three points behind the larger GLM-5.3, which scores 60, and levels it with GPT-5.6 Terra and Muse Spark 1.2.
The cost difference is stark. On the index, GLM-5.3-Flash runs at 0.09 dollars per task. The larger GLM-5.3 costs 0.68 dollars per task, making the new model roughly 7.5 times cheaper.
Artificial Analysis describes this position as the Pareto frontier of intelligence and cost. The release adds to a growing list of Chinese models that are applying heavy price pressure on Western providers.
On Z.ai’s API, the pricing is 0.15 dollars per million input tokens and 0.50 dollars per million output tokens. This is a little over ten percent of the price charged for GLM-5.3.
The model keeps pace with its larger version on agentic tasks. On the GDPval-AA v2 benchmark, it achieves an Elo score of about 1770. This matches GLM-5.3 and Grok 4.6, and trails only Claude Opus 5.
Efficiency remains a weakness. Artificial Analysis found that roughly 90 percent of the output tokens consumed went to reasoning.
Hardware independence
Before the official launch, Z.ai tested the model anonymously as “ox-alpha” on OpenCode and OpenRouter. It became the most popular model of the week during that period.
Z.ai states that all traffic for those tests ran on Chinese AI chips. SemiAnalysis reports the system served 100 trillion tokens a day.
That level of capacity was previously thought possible only for frontier labs. Z.ai claims its hardware efficiency and cost per token are on par with common Nvidia GPUs.
SemiAnalysis views this as another test of the “CUDA moat”. This follows recent results from OpenAI’s new chip.
CUDA is Nvidia’s programming layer between AI software and the graphics card. It has grown for nearly 20 years, and just about every AI framework is tuned for it.
Switching to other chips means redoing that work. Developers must reprogram compute operations, adjust memory access, and hunt down bottlenecks.
Z.ai built its own serving software on top of SGLang. It broke processing into stages that scale independently. The team says this tripled throughput over its first attempt on the same hardware.
An agent based on GLM-5.3 helped with the optimisation.
