In the past couple of days, an anonymous model named Ox Alpha has quietly appeared on the OpenRouter platform, with the codename stealth/ox-alphaThis model was launched without any company announcement or brand promotion, appearing only as an anonymous provider named “Stealth.” In Chinese-speaking circles, it is nicknamed “牛來” (Ox is Coming) as a homophone. The model supports 1M context, multimodal reasoning, and delivers excellent performance, and its free availability has sparked widespread discussion.
Ox Alpha Core Specifications: Million-Token Context and Native Multimodal
According to the OpenRouter page information, Ox Alpha launched around August 20, offering a free usage period of approximately one week, with a commitment to Zero Data Retention (no retention of user data, no use for training). Its core specifications include:
- Context window：1,048,576 tokens（約 100 萬），接近 Gemini 3.7 Flash 等級
- Maximum output: 131,072 tokens (approximately 130,000)
- Native multimodal input: Supports three formats: text, images, and video.
- Reasoning abilityReasoning mode enabled by default (reasoning effort: max), supports Tool Calling
- Service capacityThe official description states a daily processing capacity of up to 100 trillion tokens, with approximately 1.83 trillion tokens already processed.
- Speed: Measured at approximately 30-50 tokens/sec, averaging about 40t/s.
OpenRouter explicitly stated that this model is developed and operated by an anonymous third party, and OpenRouter is only responsible for routing, not the model developer.
Tokenizer, video encoder, and audio behavior triple verification highly consistent with GLM 5.3
After Ox Alpha went live, multiple independent researchers conducted systematic research on it.Technical fingerprint testing, and the results strongly point to Zhipu AI’s (Z.ai) GLM series models.
Tokenizer fingerprint matchingResearcher dax published on X the results of token count tests on 25 different prompt sets. The results show that Ox Alpha’s native token count exactly matches GLM-5.3’s, with only a fixed +75 token offset (presumably a system prompt wrapper added by OpenRouter or the provider). The tokenizer is a model’s “fingerprint”—the probability of two different base models producing identical token segmentation sequences across 25 diverse prompt sets is extremely low.
Video encoder comparisonThe same researcher further tested the token consumption of 4 control videos. Ox Alpha’s token expenditure matched GLM-5V-Turbo, Zhipu’s own multimodal model, exactly token-for-token across three independent encoding design choices: FPS-preserving frame sampling, duration scaling at approximately 147 tokens per second, and per-frame resolution scaling. The simultaneous alignment of three independent design choices is hard evidence that is difficult to explain away as “pattern-matching enthusiasm.”
Audio refusal behavior.Ox Alpha refuses audio input, with routing exactly the same as GLM-5V (no audio endpoints exposed). Xiaomi MiMo v2.5, by contrast, accepts and processes audio input. This is a binary behavioral difference that directly rules out the MiMo theory.
Furthermore, the model endpoint parameter configuration publicly available on OpenRouter shows that GLM-5.3 and Ox Alpha share the same set of parameter interfaces, and this new configuration went live just two days before Ox Alpha appeared. Their knowledge cutoffs are also identical, both extending to November 2025.
Exclude other candidates
Before the fingerprint test results were announced, the two most common guesses in the community were Xiaomi MiMo and Alibaba Qwen. However, both theories were weakened after technical verification:
- Xiaomi MiMoThe audio endpoint behavior directly contradicts (MiMo accepts audio, Ox Alpha rejects it). Previously, OpenRouter’s anonymous models Hunter Alpha and Healer Alpha were ultimately confirmed to be preview versions of MiMo, but this time the fingerprint evidence points to a different lab.
- Qwen 3.8After being included in the same round of fingerprint testing, it produced Tokenizer and encoder features that were clearly different from Ox Alpha.
Discussion on ZhihuIt was also pointed out that someone compared token counts using 25 completely different prompt sets and found that Ox Alpha’s raw token count was almost identical to GLM-5.3’s, differing by only a fixed extra 75 tokens each time. It is currently widely believed that Ox Alpha is the multimodal version of GLM-5.3 (with… DeepSeek-V4-Flash-Vision-Exp Same route)
Performance Benchmark: DeepSWE Subset Testing Delivers Impressive Results
In terms of actual performance, the community has conducted multiple tests on Ox Alpha. The most notable is the DeepSWE Coding test, which does not ask the model to solve an algorithm problem, but instead resembles a real software engineering task: given a real codebase and specific requirements, the model must understand the code on its own, identify issues, modify it, and complete the task. In a small-sample test of 10 tasks, the results are as follows, with its performance surpassing the SOTA models from both O and A:
- Ox Alpha: about 80%
- Claude Fable 5：65%
- GPT-5.6 Sol：52%
Why choose to publish anonymously?
If Ox Alpha is indeed Zhipu’s model, why hide its identity? The answer may lie in the value of the “stealth release” strategy itself.
Going live anonymously on a neutral platform allows for stress testing at scale with real production traffic before a model officially carries a lab’s name and reputation into public scrutiny. OpenRouter’s backend data shows that just two tools—Claude Code and Hermes Agent—pushed billions of tokens of traffic through Ox Alpha within a matter of days.
This practice has precedent: the anonymous preview models Hunter Alpha and Healer Alpha on OpenRouter were both eventually confirmed to be releases of Xiaomi’s MiMo. Zhipu founder Tang Jie had also previously stated that the multimodal mode would be launching soon, and about three months have passed since that statement.
The Evolutionary Path from Coding Agent to Multimodal Agent
The significance of Ox Alpha may go beyond just “GLM getting a bit stronger.” The publicly available GLM-5.3 is a pure text model, while Ox Alpha adds image and video input capabilities, paired with a million-token context, suggesting its technical path may have evolved from a Coding Agent to a multimodal agent. This means the agent can no longer only read code and operate a terminal—it can now directly “see” screenshots, “see” webpages, and “see” videos, then invoke tools and execute tasks based on what it sees. The value of this direction is far greater than a few extra points on some benchmark.
Additionally, DeepSeek has just recently added vision capabilities, and the multimodal agent roadmap for Chinese models is accelerating. If Zhipu ultimately confirms Ox Alpha’s identity, then GLM-5.3 may only be the starting point.
How do developers experience
Ox Alpha is currently available via the OpenRouter API (stealth/ox-alpha) and OpenCode Zen are available for free. Developers can quickly integrate them into tools like OpenClaw, Hermes, Cursor, and Aider through a standard OpenAI-compatible interface (just tell the Agent to add this model; testing so far shows it performs quite well). Its million-token context window allows directly inputting an entire codebase for refactoring without RAG chunking; video input can also be used to diagnose frontend interaction bugs.
Notes for this stealth model:
💰 It is free
🔑 This time, the provider does not train on your prompts or completions— OpenRouter (@OpenRouter) August 20, 2026
The free preview period is about a week, after which the pricing and official identity will be revealed, and we should have answers soon.
Source: KOCPC Chinese