GLM-5.3-Flash represents a deliberate architectural departure within the GLM-5 line, built from a freshly trained base rather than post-trained on the larger sibling's foundation. The model pairs a 320-billion-parameter mixture-of-experts design with only 18 billion parameters active per token, trained on a 30-trillion-token multimodal corpus. Its hybrid stack combines sparse and linear attention across 45 layers, with the majority using linear attention, and routes each token to 8 of 288 experts. This structure is described as reducing attention computation by roughly three times and KV-cache size by about four times relative to the text-only flagship, which is what makes the long-context serving cost practical.
As the first natively multimodal entry in the GLM-5 series, the model accepts text, image, and video input and returns text, opening workflows that earlier siblings could not handle, such as browser automation, computer-use agents, frontend visual coding, and screenshot-heavy audit tasks. Independent evaluation places it at 57 on the Artificial Analysis Intelligence Index, close to the larger flagship's 60, while remaining roughly an order of magnitude cheaper to serve. Open weights ship under the MIT license, making it the strongest openly downloadable GLM model currently available and a strong fit for high-volume agents, visual reasoning, and cost-sensitive production deployments where the flagship's extra capability is not required.
