The Model the Internet Fingerprinted as OX Alpha Is Now Officially Open Weight

Zhipu AI's GLM-5.3-Flash landed in HuggingFace Transformers v5.16.1 as official open weights — the same model architecture the developer community fingerprinted to OpenRouter's OX Alpha with high confidence in late August.

The design: a mixture-of-experts model with 320B total parameters and 18B active per forward pass. Per Zhipu AI's own release notes (not independently verified), it claims to approach Claude Opus 4.8 on coding and agentic benchmarks. Text, image, and video input; tool-calling supported.

The hardware reality: 18B active parameters sounds locally runnable, but loading the full 320B expert pool requires server-class storage and memory — you're looking at quantized GPTQ or GGUF builds before Ollama is realistic. The OX Alpha free-tier window on OpenRouter has closed; if you want to experiment, the open-weight path is now the practical one. Watch the HuggingFace model page for community quantizations.

Why it matters: If you need a strong open-weight model with native tool-calling for self-hosted inference, GLM-5.3-Flash is now a testable option — your existing MCP or function-calling wiring shouldn't need changes.

Sources: State of Open Models Summer 2026 — HuggingFace Blog · HuggingFace Transformers releases · OX Alpha mystery model reporting — The Next Web