In the large model community, it's popular to conduct "masked testing."
Recently, an anonymous model named Ox Alpha suddenly appeared on OpenRouter. Since "Ox" means "bull" in Chinese, netizens quickly gave it a more down-to-earth nickname:
The "Bull Comes" large model.
According to OpenRouter's disclosure, Ox Alpha offers a 1 million token context, supports text, image, and video inputs, can invoke tools, and is currently completely free.
Shortly after launch, the developers integrated it with the coding agent and deployed it in real code repositories for testing.
Initial test results show that this unverified "NiuLai" large model has capabilities approaching those of today's top code models.
Meanwhile, netizens have reported that an anonymous model named Korrine is being tested on Code Arena. Some speculate it is Kimi K3.1 is also sometimes referred to as Qwen and MiMo, with various claims circulating.
In August, the large model community suddenly turned into a massive guessing game.
The "Niu Lai" large model has performed outstandingly.
Ox Alpha truly garnered attention through its coding expertise.
Developer Ben Davis tested 10 tasks extracted from DeepSWE; Ox Alpha completed 8 of them, achieving an 80% pass rate. In the 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 evaluates real-world software engineering skills. The model must read a code repository, identify issues, modify the code, run tests, and then iteratively fix errors based on test failures. Compared to single-round programming problems, it more closely resembles the actual work of a coding agent.
However, the sample of 10 tasks is small. Subsequently, other developers tested on a different subset of DeepSWE and achieved results of approximately 63%. The tasks and execution configurations for the two tests were not identical, so it is not yet possible to determine Ox Alpha’s precise ranking based on these results.
However, these preliminary results indicate that this anonymous model has already demonstrated strong long-range encoding potential, approaching the capabilities of current leading models.
Who owns the "Niulai" large model?
The most widely circulated claim regarding the identity of the "Niu Lai" large model is that it is either the unreleased GLM-5.3 Flash or the multimodal version of GLM-5.3.
Someone even wrote a blog post analyzing it:
1. The strongest evidence comes from the video encoder. Across four videos with varying frame rates, durations, and resolutions, Ox Alpha consumed the exact same number of visual tokens as GLM-5V-Turbo, while MiMo, Qwen, and GLM-4.6V all showed significantly different results.
2. The text tokenizer is also highly consistent. Researchers tested 25 sets of prompts, and Ox Alpha consistently maintained a fixed difference of 75 tokens from GLM-5.3.
3. Other features also point to Zhipu. Ox Alpha refuses to process audio, just like GLM-5V’s routing method; its response style, number of Agent execution steps, and reasoning interface are also very similar to GLM. Zhipu previously used Pony Alpha to anonymously test GLM-5, demonstrating the same operational precedent.
https://ox-alpha-evidence-production.up.railway.app/
Some netizens also found clues from the conversation.
Ben Davis is 99% certain this is GLM-5.x.
These clues enhance the credibility of GLM's statement, but they are not sufficient to complete identity verification.
As of now, OpenRouter and Zhipu have not issued any public response.
Korrine's identity is even more mysterious.
The identity of the "Niulai" large model remains unclear, and another anonymous model named korrine has appeared on Code Arena.
Initially, many speculated that it was Kimi K3.1, because Kimi Before its release, K3 was believed to have been tested under the code name "kivine." Due to its structural similarity to "korrine," this led to associations.
However, the original source who first leaked the news later added that the previously reported Moonshot new model may correspond to another codename, adamant-ananke. Korrine could also originate from other Chinese teams, such as Qwen.
In the comments section, some users are also pointing to MiMo V3.
Why do large model providers prefer to "wear masks"?
Anonymous testing is becoming a crucial step before the official release of large models.
Hiding the manufacturer and model in Arena helps minimize brand bias. Users cannot see the model’s identity and must choose based solely on actual performance, resulting in battle outcomes that better reflect real-world user experience.
OpenRouter provides a different type of testing environment.
Developers will integrate the model with various coding agents, allowing it to enter real repositories, continuously invoke tools, and handle software engineering tasks lasting several hours. Whether the context remains stable, whether tool calls are reliable, and whether the model loops or veers off course during long-term tasks can all be quickly revealed under intensive use.
For model providers, this is equivalent to a public stress test. The team can observe failure cases in advance, validate the capacity of the inference service, and build authentic reputation before the official launch.
In addition, "guess the model" is increasingly becoming a marketing tactic, as identity suspense can indeed extend the discussion cycle.
Finally, returning to the model itself: if Ox Alpha is truly a flash model whose coding capabilities are already approaching the top tier, where might the upper limit be pushed by a more fully-featured version with greater resource investment?
Reference link:
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 "Machine Heart" (ID: almosthuman2014), authored by someone interested in AI.