GLM-5.3-Flash turns the anonymous Ox Alpha trial into an open-weight multimodal coding model

On August 26, 2026, Z.ai released GLM-5.3-Flash and disclosed that the model had been tested anonymously as Ox Alpha on OpenCode and OpenRouter. GLM-5.3-Flash is a natively multimodal mixture-of-experts model with 320B total parameters and 18B active parameters per token. Z.ai says it combines sparse and linear attention, supports long-context workloads up to one million tokens in its hosted path, and is available as public weights with documented SGLang, vLLM and TokenSpeed deployment paths. Cloudflare also added GLM-5.3-Flash to Workers AI on launch day. After release, Z.ai updated the model’s chat template and explicitly told users who had downloaded it before August 27 at 00:30 UTC+8 to download the corrected template again.

This is not merely a renamed GLM-5.3 checkpoint. GLM-5.3-Flash is an efficiency-oriented open model that moves multimodal coding and agent capability into a sparse architecture with far fewer active parameters than its total size suggests. For builders, that creates a new trade-off between hosted frontier models and self-controlled inference. The immediate chat-template correction adds an operational lesson: open weights make the stack inspectable and controllable, but early adopters also own model-package compatibility. A stale prompt template can change how messages, tools or reasoning instructions are serialized even when the weight files themselves are unchanged.

Z.ai says it served GLM-5.3-Flash anonymously as Ox Alpha on OpenCode and OpenRouter before release to gather real-world feedback. OpenCode reported very large usage during the trial, making the launch unusual in that the model had already been exercised at production-like scale before its public identity and weights were disclosed. Usage volume is evidence of scale and demand, not independent proof of benchmark quality.

GLM-5.3-Flash has 320B total parameters but activates 18B per token. Z.ai says the model combines sparse attention with linear attention and uses IndexPool to reduce long-context indexing overhead. Compared with GLM-5.3, the company reports materially lower attention compute and KV-cache requirements. These are vendor claims and should be treated as directional until independently reproduced under comparable serving configurations.

Z.ai has released the model weights and documents support for SGLang, vLLM and TokenSpeed. That allows builders to benchmark the model on their own repositories, agent harnesses, hardware and privacy constraints instead of relying only on the hosted coding plan. Community quantizations appeared immediately, but their quality and throughput vary by format and runtime.

Z.ai has already updated the GLM-5.3-Flash chat template after launch. In the project’s Hugging Face discussion, the Z.ai organization tells users who downloaded the model before August 27 at 00:30 UTC+8 to download the chat template again. Builders who mirrored or pinned the launch-day repository should therefore verify the template revision before debugging model behavior or comparing self-hosted results with hosted endpoints.

Cloudflare added GLM-5.3-Flash to Workers AI on August 26. That gives developers an OpenAI-compatible and Workers-native deployment path while keeping the option of self-hosting. The simultaneous open-weight and managed-inference availability lowers the cost of comparing control, latency and operational overhead.