For a week an anonymous model called ox-alpha sat at the top of OpenRouter and OpenCode with nobody knowing who made it. On 26 August 2026 Z.ai (Zhipu AI) revealed it as GLM 5.3 Flash and open-sourced the weights under MIT the same day. Here is what it actually is, what the benchmarks support, and how to run it on an OpenClaw agent.
What GLM 5.3 Flash Is
GLM 5.3 Flash is a mixture-of-experts model with 320 billion total parameters and about 18 billion active per token. It is the first natively multimodal model in the GLM-5 series — text, images and video go in, text comes out — and unlike GLM 5.3, which was a post-training re-tune of GLM 5.2's base, it starts from a newly trained base model.
Despite the name, it is not a distilled small sibling of GLM 5.3. It is a different architecture aimed at a different goal: keep frontier-level capability while cutting the cost of serving it. Z.ai reports that against GLM 5.3 it reduces attention compute by 3.01x and KV cache size by 4.44x.
Hybrid sparse + linear attention — linear attention models local dependencies, a lightweight sparse indexer retrieves global context. Z.ai calls it the first open-source frontier model to combine the two.
