요약
Z.ai’s GLM-5.3-Flash, previously tested anonymously as Ox Alpha on OpenRouter, combines near-frontier coding and agentic performance with open weights, local deployment options, and exceptionally low inference costs. Matthew Berman highlights the model’s effi…
본문
Pinned postOpen weights with surprising performance
GLM-5.3-Flash is a mixture-of-experts model with 320 billion total parameters and 18 billion active parameters. Berman says it substantially improves on GLM-5.2 and approaches leading proprietary systems on coding, agentic tasks, and practical knowledge work, although it remains below much larger frontier models.
The cost advantage is the main attraction
The model reportedly costs about 9 cents per completed Artificial Analysis Intelligence Index task, compared with several dollars for some leading proprietary models. Its million-token context window, 131,000-token maximum output, and very low cached-input pricing further strengthen its appeal. However, Berman notes that GLM-5.3-Flash can use far more output tokens than some cheaper alternatives.
Read 4sysops without ads and access many member features.Chinese accelerators become part of the story
Berman cites a report that GLM-5.3-Flash processed roughly 100 trillion tokens per day on Chinese AI chips without Nvidia GPUs. He presents this as evidence that China is co-designing models, accelerators, interconnects, and serving software to achieve competitive inference economics on domestic infrastructure.
Local control and practical testing
Because the weights are available, organizations can customize, fine-tune, host, or serve the model through compatible APIs such as OpenRouter and Z.ai. Berman tested it on interactive Rubik’s Cube simulations, website generation, and a branded data-center presentation. He found it capable overall, with particularly strong results on some web designs and the presentation, while GPT-5.6 Sol produced more detailed visuals in a 3D scene.
Read 4sysops without ads and access many member features.