The Mysterious “Niu Lai” (Ox Alpha) Model: Free, High‑Performance Coding AI and Its Possible Origin

Ox Alpha, an anonymous large‑language model that appeared on OpenRouter, is nicknamed “Niu Lai” (meaning “cow arrives”). According to the platform, it supports a 1 million‑token context window, accepts text, image and video inputs, can invoke tools, and is currently free.

Developers quickly integrated Ox Alpha into coding agents and ran it against real code repositories. Using the DeepSWE benchmark, which evaluates genuine software‑engineering ability by requiring models to read repositories, locate bugs, modify code, run tests and iteratively fix failures, Ox Alpha completed 8 out of 10 tasks, achieving an 80% success rate. In the same benchmark, Fable 5 Max scored 65%, GLM‑5.3 Max and Grok 4.6 xhigh each scored 62%, and GPT‑5.6 Sol Max reached 52%.

A second set of developers tested Ox Alpha on a different DeepSWE subset and reported a roughly 63% success rate. Because the two test sets and runtime configurations differ, the exact ranking of Ox Alpha among top models remains uncertain.

The article then examines the model’s likely identity. The most circulated hypothesis is that Ox Alpha is the unreleased multimodal version of Zhipu’s GLM‑5.3 Flash. Supporting evidence includes:

Video encoder token consumption matches that of GLM‑5V‑Turbo across four videos of varying frame rates, durations and resolutions, while other candidates (MiMo, Qwen, GLM‑4.6V) differ.

Text tokenizer behavior shows a constant 75‑token offset compared with GLM‑5.3 across 25 prompt variations.

Ox Alpha refuses audio input and follows the same routing logic as GLM‑5V; its response style, number of Agent steps, and inference interface closely resemble GLM.

Zhipu previously conducted anonymous tests of GLM‑5 using the Pony Alpha alias, establishing a precedent.

Ben Davis, who performed the DeepSWE tests, estimates a 99% certainty that Ox Alpha corresponds to a GLM‑5.x model, though no official confirmation has been provided by OpenRouter or Zhipu.

Simultaneously, another anonymous model named “korrine” surfaced on Code Arena. Initial speculation linked it to Kimi K3.1 (due to the earlier “kivine” codename), while later hints suggested it might be a Moonshot model codenamed “adamant‑ananke” or a Qwen variant. Some commenters even associated it with MiMo V3, but its true provenance remains unresolved.

The article explains why model vendors favor “masked” testing. By hiding the manufacturer and model name in arenas, brand bias is reduced, forcing users to judge purely on output quality. This approach also serves as a public stress test, revealing failures in long‑running, tool‑heavy coding tasks and allowing teams to observe performance limits before official release. Moreover, the mystery itself fuels community discussion and serves as a marketing tactic.

Finally, the piece speculates on the future ceiling of Ox Alpha’s capabilities. If the current version already rivals leading code models, a full‑featured release could push performance even higher.

References: 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