It's trendy to test large models with 'anonymous accounts'.
Recently, an anonymous model named Ox Alpha suddenly appeared on OpenRouter. Ox means 'ox' or 'cow', and domestic netizens quickly gave it a more relatable nickname:
"Ox Comes" large model.
According to information disclosed by OpenRouter, Ox Alpha has a 1 million token context window, supports text, image, and video input, can call tools, and is currently completely free.
Shortly after launch, developers integrated it into a coding agent and threw it into a real code repository for testing.
Preliminary test results show that this mysterious "Ox Comes" large model's capabilities are already approaching current top-tier code models.
Meanwhile, netizens revealed that an anonymous model named korrine is being tested on Code Arena. Some speculate it's Kimi K3.1, while others point to Qwen and MiMo, with varied guesses.
In August, the large model circle has suddenly turned into a large-scale guessing game.
"Ox Comes" Model Performs Remarkably
What truly drew attention to Ox Alpha was its coding capability.
Developer Ben Davis selected 10 tasks from DeepSWE for testing, and Ox Alpha completed 8 of them, achieving an 80% pass rate. In his published comparison results, Fable 5 Max scored 65%, GLM-5.3 Max and Grok 4.6 xhigh both scored 62%, and GPT-5.6 Sol Max scored 52%.
DeepSWE examines real-world software engineering ability. The model needs to read code repositories, locate problems, modify code, run tests, and continue fixing based on error reports. Compared to single-round coding problems, it's closer to the actual work of a coding Agent.
However, 10 tasks is a very small sample. Subsequently, other developers tested on another subset of DeepSWE, reporting a result of about 63%. The task scopes and execution configurations of the two tests were not identical, making it impossible to definitively rank Ox Alpha based on this alone.
Nevertheless, these results preliminarily show that this anonymous model has demonstrated strong potential for long-context coding, with capabilities approaching current leading models.
Who Created the "Ox Comes" Large Model?
The most widespread speculation about the identity of the "Ox Comes" model is that it is an unreleased GLM-5.3 Flash from Zhipu AI, or a multimodal version of GLM-5.3.
Someone even wrote a blog to analyze this:
1. The strongest evidence comes from the video encoder. For four videos with different frame rates, durations, and resolutions, the visual tokens consumed by Ox Alpha matched exactly with GLM-5V-Turbo. MiMo, Qwen, and GLM-4.6V all showed significantly different results.
2. The text tokenizer also shows a high degree of alignment. The researcher tested 25 sets of prompts; the token count between Ox Alpha and GLM-5.3 consistently maintained a fixed difference of 75 tokens.
3. Other features also point to Zhipu AI. Ox Alpha refuses to process audio, which matches the routing method of GLM-5V; its answer style, the number of Agent execution steps, and the inference interface are also very similar to GLM. Zhipu AI previously used Pony Alpha for anonymous testing of GLM-5, establishing a precedent for this practice.
https://ox-alpha-evidence-production.up.railway.app/
Other netizens have also found clues in conversations.
Ben Davis believes he is 99% certain this is GLM-5.x.
These clues increase the credibility of the GLM theory, but are still insufficient for definitive identity confirmation.
As of now, neither OpenRouter nor Zhipu AI have publicly responded.
korrine's Identity is Even More Mysterious
While the identity of the "Ox Comes" model remains unclear, another anonymous model named korrine has appeared on Code Arena.
Initially, many speculated it was Kimi K3.1, because before the release of Kimi K3, it was believed to have been tested under the codename kivine. The similar structure of kivine and korrine sparked this association.
However, the original source of the rumor later added that the previously learned about new Moonshot model might correspond to another codename, adamant-ananke. korrine could also come from other Chinese teams like Qwen.
In the comments, some also pointed to MiMo V3.
Why Do Large Model Companies Like 'Testing in Disguise'?
Anonymous testing is becoming an important step before the official release of large models.
Hiding the manufacturer and model name in the Arena can minimize preconceptions brought by branding. Users cannot see the model's identity and can only choose based on actual outputs. The accumulated battle results are also closer to the real user experience.
OpenRouter provides a different kind of testing environment.
Developers integrate the model into various coding Agents, letting it enter real repositories, continuously call tools, and handle software engineering tasks lasting several hours. Issues like context stability, tool calling reliability, and whether the model gets stuck in loops or goes off track during long tasks can be quickly exposed under intense use.
For model developers, this is akin to a public stress test. Teams can observe failure cases in advance, verify the capacity of their inference services, and also accumulate real-world reputation before official launch.
Moreover, "guessing the model" is increasingly becoming a marketing tactic; the suspense over identity can indeed prolong the discussion cycle.
Finally, back to the model itself. If Ox Alpha is truly a Flash model and its coding ability is already approaching top-tier levels, where will the ceiling be pushed by the more resource-intensive, more capable full version?
Reference Links:
https://x.com/Adidotdev/status/2090833298713096241
https://x.com/davis7/status/2090669483740279155?s=20
https://x.com/davis7/status/2090655207831298095?s=20
https://x.com/MaxForAI/status/2090783750217162788
This article is from the WeChat public account "Almost Human" (ID: almosthuman2014), author: Following AI
