The Beijing lab told Bloomberg it built the anonymous coding model and planned to release its weights on August 26.
By RuntimeWire Staff · Published
Primary source: Bloomberg Technology
Why it matters #
Publishing Ox Alpha's weights would let developers move beyond the OpenRouter preview and independently test its architecture, performance, hardware requirements and safety behavior. The accompanying license will determine whether that scrutiny can translate into commercial modification and self-hosting.
Z.AI, the Beijing AI lab co-founded and led by Zhang Peng, told Bloomberg on August 26 that the anonymous Ox Alpha model is a new iteration of its GLM series. The company said it planned to release the model's weights that night.
That disclosure gives developers a concrete way to assess a model that OpenRouter listed as a free preview from an anonymous third-party provider. A downloadable checkpoint could allow direct testing, private deployment and closer examination of the model's architecture, subject to its hardware requirements and license.
Ox Alpha appeared around August 20 on OpenRouter, where it was presented as a reasoning model for coding, sustained agent work and production workloads. TechCrunch reported that access was free and that the developer and operator had chosen to remain anonymous during the preview. Bloomberg later reported that the model had reached the top of online usage charts.
From fingerprint to confirmation
Model-forensics researchers had previously linked the preview to Z.AI's GLM family through tokenizer, multimodal and API-behavior fingerprints, but Bloomberg's report supplies the first company confirmation in the provided reporting. That earlier work included a public GitHub investigation and an independent CTGT study. The findings supplied background for the attribution; Z.AI's statement to Bloomberg identified the model's maker.
What the planned weights would change
A downloadable checkpoint would move evaluation beyond an endpoint controlled by the provider. Developers could measure coding and agent performance on their own workloads, compare inference systems and determine the memory and compute needed to run the model. Model files and configuration data could also clarify which GLM architecture sits behind the preview.
The release terms will matter as much as the files. Bloomberg did not specify an Ox Alpha license, parameter count, training-compute figure or benchmark methodology. A permissive license could support commercial modification and self-hosting, while narrower terms could restrict how developers deploy or adapt the model.
Self-hosting would also change the data-handling calculation by keeping sensitive workloads outside the preview endpoint. Organizations able to run the model themselves under its eventual license could test sensitive code without sending those workloads through the hosted service.
Bloomberg described Ox Alpha as a DeepSeek rival, but the available reporting does not include reproducible benchmark runs establishing parity. Weight access would let independent researchers test that comparison across coding tasks, long-context work and agent workflows rather than relying on performance through a single hosted service.
Zhang's open-weight distribution bet
Zhang worked at Tsinghua University from August 2005 through December 2020 before taking charge of the commercial lab that grew from its research community. Z.AI's listing prospectus identifies him as co-founder, chief executive, general manager and executive director, responsible for research, business development and daily operations.
In an April founder's statement, Zhang traced the company's starting point to the team's late-2018 interest in dual-system theories of human cognition. Z.AI was founded in 2019 and went on to develop the GLM pre-training framework and a succession of language, coding and multimodal models.
The company is financing that work through steep losses. Its prospectus reported RMB312.4 million in 2024 revenue and an approximately RMB3.0 billion net loss, with research spending reaching RMB2.20 billion. The US Commerce Department also added multiple Zhipu AI entities to the Entity List in January 2025, imposing export-license requirements with a presumption of denial for covered items.
Ox Alpha's anonymous preview generated usage before Z.AI attached its name. Publishing the weights would expose the model to a harder test: whether developers can reproduce its hosted performance, run it economically and obtain enough control under the license to build on it.