The mysterious "Niu Lai" large model has gone viral across the internet, and it is available for free for a limited time.

The practice of "testing under a pseudonym" is widely popular in the large model circle.

Recently, an anonymous model named Ox Alpha suddenly appeared on OpenRouter. "Ox" itself means cattle, and domestic Chinese netizens quickly gave it a more down-to-earth name:

The "Niu Lai" Large Model.

According to the 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 of charge.

Shortly after its launch, developers connected it to a coding Agent and tested it in a real code repository.

Preliminary test results show that this unidentified "Niu Lai" large model has capabilities close to the current top-tier code models.

Meanwhile, some netizens broke the news that an anonymous model named korrine is being tested on Code Arena. Some people guess it is Kimi K3.1, others point to Qwen and MiMo, and there are all kinds of speculations.

The large model circle in August has suddenly turned into a huge guessing game.

The "Niu Lai" Large Model Delivers Impressive Performance

Ox Alpha has really drawn attention for its outstanding coding capability.

Developer Ben Davis selected 10 tasks from DeepSWE for testing, and Ox Alpha completed 8 of them, with a pass rate of 80%. In the published comparison results, Fable 5 Max scored 65%, both GLM-5.3 Max and Grok 4.6 xhigh scored 62%, and GPT-5.6 Sol Max scored 52%.

DeepSWE assesses real software engineering capabilities. The model needs to read code repositories, locate problems, modify code, run tests, and continue to fix issues according to error reports. Compared with single-round programming problems, it is closer to the actual work of coding Agents.

However, the sample size of 10 tasks is very small. Later, other developers tested it on another subset of DeepSWE, and the resulting pass rate was about 63%. The task scope and operation configuration of the two tests are not completely identical, so it is impossible to determine the exact ranking of Ox Alpha for the time being.

Nevertheless, these results initially show that this anonymous model has demonstrated strong long-horizon coding potential, with capabilities close to the current leading models.

Who Is the Developer of the "Niu Lai" Large Model?

As for the identity of the "Niu Lai" large model, the most widely spread claim at present is that it is the unreleased GLM-5.3 Flash or the multimodal version of GLM-5.3 from Zhipu AI.

Some people even wrote a dedicated blog for analysis:

1. The strongest evidence comes from the video encoder. In four videos with different frame rates, durations and resolutions, the visual tokens consumed by Ox Alpha are completely consistent with those of GLM-5V-Turbo. MiMo, Qwen and GLM-4.6V all show significantly different results.

2. The text tokenizer is also highly consistent. Researchers tested 25 sets of prompts, and the number of tokens of Ox Alpha and GLM-5.3 always maintained a fixed difference of 75 tokens.

3. Other features also point to Zhipu AI. Ox Alpha refuses to process audio, which is the same as the routing method of GLM-5V; its answer style, Agent execution steps and inference interface are also very close to GLM. Zhipu AI previously used Pony Alpha to anonymously test GLM-5, so it has the precedent of doing the same operation.

https://ox-alpha-evidence-production.up.railway.app/

Some other netizens also found clues from the conversation records.

Ben Davis believes that it is 99% certain that this is GLM-5.x.

These clues increase the credibility of the "GLM theory", but are not sufficient to confirm the identity of the model.

As of now, neither OpenRouter nor Zhipu AI has made a public response.

The Identity of korrine Is Even More Confusing

While the identity of the "Niu Lai" large model is still unclear, another anonymous model named korrine has appeared on Code Arena.

At first, many people guessed that it was Kimi K3.1, because before Kimi K3 was released, it was known to be tested under the code name kivine. The similar structure of the names kivine and korrine led to such association.

However, the whistleblower who first spread the news later added that the previously known new Moonshot model may correspond to another code name adamant-ananke. korrine may also come from other Chinese teams such as Qwen.

In the comment section, some people also pointed it to MiMo V3.

Why Do Large Model Vendors Prefer "Testing Under a Pseudonym"?

Anonymous testing is becoming an important step before the official release of large models.

Hiding the vendor and model name in the Arena can minimize the preconceived impression brought by the brand. Users cannot see the identity of the model, and can only make choices based on the actual output. The accumulated battle results are also closer to the real user experience.

OpenRouter provides another type of testing environment.

Developers will connect the model to various coding Agents, let it enter real repositories, continuously call tools, and process software engineering tasks that last for hours. Whether the context can remain stable, whether the tool calls are reliable, and whether the model will loop or deviate in long-horizon tasks can be quickly exposed in high-intensity usage scenarios.

For model vendors, this is equivalent to a public stress test. Teams can observe failure cases in advance, verify the carrying capacity of the inference service, and accumulate real public praise before the official release.

In addition, "guessing the model" has increasingly become a marketing method now, and the suspense of identity can indeed extend the discussion cycle.

Finally, back to the model itself. If Ox Alpha is really a Flash model and its coding capability is already close to the leading level, where will the full version with higher resource investment and more complete capabilities push the upper limit to?

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 Official Account "Machine Heart" (ID: almosthuman2014), the author is an AI-focused content creator, and 36Kr publishes this content with authorization.