Z.ai GLM-5.3-Flash: 3.3x Faster, Cheaper AI Model

Z.ai has unveiled its new GLM-5.3-Flash model, which runs 3.3 times faster on a single workstation. This model, previously tested anonymously as Ox Alpha, is now available with an MIT license on Hugging Face. Here's the thing: GLM-5.3-Flash is a 320-billion-parameter model, but it's designed so fewer parameters are active at any given time. This architecture significantly reduces computing costs per step. What's interesting is it's the first natively multimodal model in the GLM-5 series, accepting text, images, and video as input to return text. The model card shows a large context window of over one million tokens. This is made possible because only a fraction of its layers store the full conversation history. This design helps keep long-context interactions less resource-intensive. Z.ai states that attention computation is reduced by about 3.01 times, and the average KV cache per layer falls by about 4.44 times compared to GLM-5.3. The bottom line: GLM-5.3-Flash beats Claude Opus 4.8 on key agent benchmarks while costing significantly less, putting pressure on pricing in the AI industry.