Skip to content
NewsModel release

Z.ai reveals Ox Alpha as GLM-5.3-Flash and releases MIT weights on Hugging Face

· by Pondero Newsdesk

The short version

Zhipu AI confirmed August 26 that mystery model Ox Alpha is GLM-5.3-Flash, a 320B MoE with 18B active parameters, native multimodal input, MIT weights on Hugging Face, and a 1M-token context window priced at $0.075 per million input tokens through September 9.

Z.ai reveals Ox Alpha as GLM-5.3-Flash and releases MIT weights on Hugging Face

Z.ai confirmed on August 26 that Ox Alpha, the anonymous model we tracked when it first appeared on OpenRouter, is GLM-5.3-Flash: a 320-billion-parameter mixture-of-experts model published under an MIT license on Hugging Face the same day, per Z.ai's official blog.

What Z.ai released

GLM-5.3-Flash is the first natively multimodal release in the GLM-5 series. It runs 18 billion active parameters out of 320 billion total (down from 32 billion active in GLM-4.5) and accepts text, images, and video in a 1,048,576-token context window with a 131,072-token output ceiling, per Z.ai. Training used a 30-trillion-token multimodal corpus.

A hybrid linear-plus-sparse attention design called IndexPool cuts per-token attention compute by 3x and KV cache size by 4.4x compared with GLM-5.3. On Z.ai's internal evaluations, the model scored 63.4 on DeepSWE v1.1 versus 46.2 for GLM-5.2, and 48.8 on AutomationBench versus 26.2. Z.ai also reports a score of 57 on the Artificial Analysis Intelligence Index v4.1.1 at $0.045 per task at discounted pricing, per the announcement. These results use Z.ai's own evaluation harnesses; independent replication is not yet available.

The Chinese-GPU infrastructure claim

During the six anonymous preview days, Z.ai says it served all traffic on Chinese-made AI chips, with the model reaching the top spot on OpenCode and OpenRouter by usage, per Z.ai. The company describes a custom SGLang-based inference stack that achieves "hardware efficiency and per-token cost comparable to mainstream NVIDIA GPUs" on domestic accelerators. No specific chip vendor has been named and the performance comparison has not been published with methodology. If subsequent audits confirm the claim, it would update assumptions about Chinese inference infrastructure capability at frontier scale.

Pricing and migration

Standard pricing is $0.15 per million input tokens and $0.50 per million output tokens, with a 50% launch discount through September 9 dropping those to $0.075 input, $0.015 cached input, and $0.25 output, per Z.ai. The weights on Hugging Face support local deployment via SGLang, vLLM, and TokenSpeed.

The stealth/ox-alpha slug is gone from OpenRouter's active catalog, per CellCog's post-reveal tracking. Operators who built on the free preview endpoint need to migrate to z-ai/glm-5.3-flash or a self-hosted deployment before September 9.

Why it matters

MIT licensing is the concrete differentiator. Competing frontier-class models with comparable benchmark scores do not ship with open weights and commercial-use rights. Operators running fine-tuning workflows, air-gapped deployments, or production stacks with IP constraints have a new option that did not exist last week at this price tier.

The weights also make independent benchmarking possible for the first time. Before August 26, every performance number for GLM-5.3-Flash came from Z.ai's own harnesses applied to the anonymous endpoint. Researchers can now replicate conditions directly and test whether the coding and vision scores hold outside Z.ai's infrastructure. Those results will settle what the stealth window left open.

What to watch next

External fine-tuning runs from the research community should appear within days and give the first independent signal on the coding scores. The Chinese GPU claim is the larger outstanding question: Z.ai has not named the chip vendor or published the methodology behind the NVIDIA efficiency comparison.

Sources