Z.ai Confirms Ox Alpha Is New GLM Model, Weights Released

Z.ai Unveils Ox Alpha: A New GLM Model with Open Weights

After days of intense speculation, Chinese AI lab Z.ai has officially confirmed that the mysterious Ox Alpha model, which took the developer community by storm, is a new iteration of its GLM series. The company announced that it has released the model's weights on Hugging Face, allowing developers worldwide to download, modify, and build upon it. This move marks a significant escalation in the competitive landscape of open-weight AI models.

Ox Alpha first appeared on August 20 on OpenRouter, an AI model routing platform, as an anonymous "stealth model." It was offered completely free with near-unlimited usage, and its performance quickly caught the attention of developers and tech leaders alike. Stripe CEO Patrick Collison publicly praised the model on X, calling it "very impressive," while early benchmarks suggested it could rival leading models from US companies like OpenAI and Anthropic.

What We Know About Ox Alpha

Z.ai, also known as Zhipu, revealed that Ox Alpha is an anonymous preview of its upcoming GLM-5.3-Flash model. The company used the stealth approach to gather user feedback on OpenRouter and OpenCode before the official release. This strategy mirrors a previous test where Z.ai ran GLM-5 under the name "Pony Alpha" ahead of its launch.

The model is described as a multimodal reasoning system designed for coding, sustained agentic work, and production workloads. It supports text, images, and video inputs, and boasts a context window of up to 1 million tokens—a significant advantage for complex, long-horizon tasks. OpenRouter characterized it as ideal for long-horizon software engineering, complex reasoning, and workflows that combine text with visual context.

How Developers Identified Z.ai

The identification of Z.ai as the creator of Ox Alpha was a collaborative effort by the developer community. A developer known as unclecode, who created the Crawl4AI open-source crawler, built a browser tool called modelprint that fingerprints anonymous API endpoints. This tool sends nine infrastructure probes to an endpoint and matches responses against known models. Ox Alpha lined up with GLM-5.3 on six of the nine probes, including all four normalized tokenizer counts—strong evidence of shared infrastructure.

While some analysts initially pointed to other labs like Xiaomi's MiMo team or Microsoft's MAI family, the tokenizer analysis and other technical fingerprints made Z.ai the leading candidate. The company had previously shipped GLM-5.3 on August 14, just six days before Ox Alpha appeared, further supporting the connection.

Why This Matters for the AI Industry

The release of Ox Alpha's weights is a major development in the ongoing competition between US and Chinese AI labs. Z.ai's move to open-source a model that rivals frontier systems at zero cost threatens the business models of companies like OpenAI and Anthropic, which rely on paid access to their most advanced models. This trend is not isolated—Chinese labs like DeepSeek and Moonshot AI have also released open-weight models that challenge US dominance at a fraction of the cost.

Moonshot's Kimi K3, a 2.8 trillion-parameter open-weight model released in July, quickly drew attention for its performance and lower price. The success of such models underscores a broader shift toward accessible, high-performance AI that could democratize access to advanced capabilities but also raise concerns about the sustainability of proprietary AI businesses.

Technical Specs and Performance

While detailed benchmark scores for Ox Alpha have not been officially published, early testing suggests it performs on par with leading models. One developer, Davis, ran the full DeepSWE set and found that Ox Alpha came out roughly level with GPT-5.6-sol mid, rather than significantly outperforming it as some initial claims suggested. This indicates that while Ox Alpha is a strong contender, it may not be a clear winner across all tasks.

The model's free access period, which ran through August 24 on OpenRouter and up to a week from launch on OpenCode, generated massive interest. OpenCode advertised capacity of 100 trillion tokens per day—roughly 100 times the number of AI tokens Visa processes in a month—highlighting the scale of the offering.

Z.ai's Growing Influence

Z.ai has been steadily building its reputation in the AI space. Earlier this month, the company released GLM-5.3, which rivals Anthropic's Fable 5 on certain benchmarks. The company also gained attention when Hugging Face used GLM models to defend against an attack from OpenAI agents. With the release of Ox Alpha's weights, Z.ai is positioning itself as a major player in the open-weight AI ecosystem.

The decision to open-source the model aligns with a broader trend in China, where many AI labs prioritize transparency and community-driven development. This approach has won favor among developers who value the ability to customize and deploy models on their own infrastructure, free from vendor lock-in.

The Road Ahead

As Ox Alpha's weights become available, expect a flurry of activity from developers eager to fine-tune the model for specific use cases. The model's multimodal capabilities and large context window make it particularly attractive for applications in software engineering, data analysis, and agentic workflows. However, the open-weight nature also raises questions about potential misuse, as with any powerful AI model.

For now, the AI community is celebrating a new milestone in open-weight AI development. Z.ai's confirmation and release of Ox Alpha's weights not only solves the mystery but also sets the stage for a new wave of innovation. The question remains: How will US companies respond to the growing threat of cheap, capable models from China? Only time will tell.