Z.ai opens GLM-5.3-Flash
Z.ai has released MIT-licensed weights for GLM-5.3-Flash and confirmed that it was ox-alpha, the anonymous model tested on OpenCode and OpenRouter last week. It is a native multimodal model with 320 billion parameters in total and 18 billion active for each token.
Developers can now inspect and modify the weights, then deploy the coding and long-running agent model with SGLang, vLLM or TokenSpeed. The concrete gain is control: teams can evaluate it inside their own infrastructure instead of using only Z.ai’s hosted service.
Open weights do not mean laptop-ready. The official repository has 62 weight shards totaling about 328 GB. Z.ai did not release the training data or a full training pipeline. Its benchmark, adoption and price comparisons are not independently verified. These include claims that the model costs one-tenth of GLM-5.2 and was the most-used model in the ox-alpha test.
1August 26, 2026 141 1
