Anonymous large model Ox Alpha tops the charts.

"Ox Coming" Makes a Splash in the AI Circle!

On August 20th, a large model with the mysterious codename Ox Alpha quietly appeared on the model aggregation platform OpenRouter and the programming tool OpenCode. There was no press conference, no technical report—just a single line in the official introduction:

A cutting-edge model designed for efficient programming, long-duration agent tasks, and real-world production environments.

Because "Ox" means "bull" in English, and coincidentally aligns with the viral animation film "Ox Coming" sweeping the internet, Chinese netizens quickly gave it a down-to-earth nickname: the "Ox Coming Large Model."

But soon, netizens discovered that Ox Alpha wasn't just impressive in name—it was powerful in reality.

Leveraging free access, ultra-long context, and outstanding programming capabilities, Ox Alpha surged to the top of OpenRouter's usage rankings within just one day of launch, setting a new single-day model usage record for the platform. Even the CEO of payment giant Stripe personally tried it out, leaving four words: "Pretty damn impressive."

But what piqued curiosity even more than its performance was Ox Alpha's origin. Overnight, the AI community transformed into detectives: Was this another hidden move by Zhipu, or had Google unleashed a major innovation?

Where's the "Bull"? Free and Formidable

Ox Alpha's rapid rise to the top within a day was first and foremost due to an attractive combination: free access, ultra-long context, and capabilities tailored for programming and agent tasks.

According to information disclosed by OpenRouter, Ox Alpha boasts an ultra-long context window of approximately 1.0486 million tokens, with a maximum output capacity of 131,000 tokens in a single go. It supports text, image, and video inputs while also enabling external tool calls.

How impressive is this? Simply put, it can ingest content equivalent to a large code repository or hundreds of pages of documentation in a single pass. For programming agents requiring long-duration operation, this means the model won't frequently forget previous context, nor will developers need to constantly curate relevant information.

And this precisely aligns with the scenarios Ox Alpha targets.

OpenRouter officially positions it for programming, continuous agent tasks, and production environments. OpenCode quickly integrated Ox Alpha after its launch, declaring its service could handle 100 trillion token calls per day.

In essence, Ox Alpha was designed from the ground up to work, not just chat.

Next came the wave of real-world tests that propelled the "Ox Coming Large Model" into the spotlight.

Independent researcher Ben Davis selected 10 real-world tasks from the software engineering benchmark DeepSWE. Ox Alpha passed 8 of them, achieving an 80% pass rate.

Among the same set of tasks, Claude Fable 5 had a 65% pass rate, while GLM-5.3 and Grok 4.6 both scored 62%, and GPT-5.6 Sol managed only 52%. The narrative of an "anonymous model outperforming flagship models" quickly spread across the internet.

Subsequently, Wenqi Huang, lead author of DeepSWE, re-ran the full 113 official tests. Ox Alpha's pass rate dropped to around 63%, placing it just slightly behind top closed-source models like Grok 4.6.

What proved even more convincing than benchmark scores were endorsements from Silicon Valley luminaries.

After personally testing Ox Alpha, the CEO of payment giant Stripe left the comment "Pretty damn impressive"; the founder of agent product Hermes Agent stated that Ox Alpha outperformed their internal evaluation standards; CEOs of companies like OpenRouter and T3 Chat also joined the chorus of endorsements.

Developers even conducted some playful tests. Some had Ox Alpha control a game agent in "Frogger," where the model not only completed the task but also added new game elements on its own. Others tasked it with creating personal websites, generating multiple design styles in one go.

While such examples can't replace rigorous professional evaluations, they offer intuitive demonstrations of its multimodal understanding, tool invocation, and code execution capabilities.

As of August 24th, Ox Alpha remained firmly at the top of OpenRouter's usage rankings.

Image Source: OpenRouter

In other words, Ox Alpha didn't just go viral based on a single impressive ranking number—it combined free access, million-token-level context, programming capabilities, and agent execution abilities into a compelling package.

The Internet Investigates the Identity of the "Ox Coming Large Model"

Since the "Ox Coming Large Model" launched anonymously, its identity naturally became the biggest suspense .

Initially, speculations ran wild. Some guessed it was Xiaomi's MiMo, given that Xiaomi's CFO had publicly stated a new MiMo was in training and had a history of anonymous testing under "Hunter Alpha." Others suspected Tencent's Hunyuan, reasoning that few manufacturers could support such large-scale free access.

As discussions heated up, netizens quickly entered "evidence-gathering" mode.

Developers began examining Ox Alpha's various "engineering fingerprints" left during text, image, video processing, and API calls—subtle implementation habits unique to different models. These included how text was segmented into tokens, how APIs handled parameters, error messages returned, and input length calculations for images and videos.

Clues from Ox Alpha's tests gradually pointed toward Zhipu's GLM.

Twenty-five prompts were tested on both Ox Alpha and Zhipu's GLM-5.3, revealing highly consistent token segmentation results. Subsequent video input tests showed Ox Alpha's processing patterns closely resembled Zhipu's publicly disclosed GLM-5V-Turbo, while results differed significantly when tested against models like Xiaomi's or Tongyi Qianwen.

Follow-up black-box testing further reinforced this conclusion. Researchers compared tokenizers, API parameters, error messages, and multi-round tool invocations, finding strong alignment between Ox Alpha and Zhipu's GLM-5 series across multiple low-level features. One recent test even achieved near-perfect matches across 25 tokenizer tests.

Adding to the intrigue, Zhipu's previous anonymous model was codenamed Pony ( pony ), while this one switched to Ox ( ox ). Netizens joked: "Whether considering zodiac classifications or the metaphor of 'oxen and horses,' this combination makes perfect sense."

Just as rumors about Zhipu spread widely, Google joined the fray with playful hints.

Researchers from Google DeepMind mentioned "Gemini" on social media, leaving cryptic comments like, "What if Ox Alpha turns out to be an old friend we've encountered along the way?" The vague wording neither confirmed nor denied involvement, adding another layer of suspense to the guessing game.

Image Source: X

As of now, Ox Alpha remains unclaimed by any party. The mystery will likely only be solved after the free testing period ends.

Interestingly, this isn't the first time an anonymous large model "unmasking" game has played out.

Large Model Vendors Embrace Anonymous Launches

Ox Alpha marks the fifth anonymous model (Stealth Model) launched on OpenRouter in recent years.

Reviewing previous anonymous models' unveilings reveals this has become a new playbook for model vendors testing new innovations.

Pony Alpha, launched in February, surpassed 40 billion token calls on its first day and was confirmed by Zhipu five days later as part of their GLM series.

Hunter Alpha, launched on March 11th with trillion-parameter scale, was initially speculated to be DeepSeek's new model but was eventually claimed by Xiaomi as MiMo-V2-Pro.

Later, Elephant Alpha in April was revealed by Ant Group's Inclusion AI about two weeks post-launch, while Owl Alpha in late April wasn't confirmed as Meituan's LongCat-2.0 until June 30th.

Notably, all four previously unveiled anonymous models came from Chinese teams. Why have model vendors increasingly chosen anonymous launches?

First, anonymity minimizes preconceptions driven by brand recognition. Developers evaluate models solely based on performance, providing authentic feedback that serves as cost-effective market testing for vendors.

Compared to vendor-designed benchmarks, deploying models in real code repositories and agent frameworks—where developers execute dozens or even hundreds of steps over hours of complex tasks—better exposes issues like context stability and tool reliability. This essentially functions as large-scale "stress testing."

This approach also offers strategic flexibility. If a model performs exceptionally well, vendors can reveal its identity at peak Heat , transforming test results into a formal launch. If performance falls short, testing can conclude quietly without brand damage.

For Chinese model vendors expanding overseas, anonymity offers additional advantages.

While products like DeepSeek and Tongyi have gained recognition among overseas developers, many domestic models still need time to build awareness.

Entering overseas developer communities with anonymous models first allows gathering authentic feedback. Revealing the identity after achieving strong results effectively splits a single launch into two waves of publicity, sustaining momentum longer.

Thus, anonymous models often create a unique propagation effect: developers handle testing, netizens fuel discussions, while vendors observe from behind the scenes whether the model truly resonates with users.

Following previous anonymous models' release patterns, Ox Alpha's free testing window will likely close around August 27th. Will it prove to be Zhipu's hidden innovation, Google's unreleased Gemini variant, or perhaps come from an unexpected player? The answer may arrive soon.

References:

"Mysterious 'Ox Coming' Model Goes Viral, Steals DeepSeek's Spotlight: How Good Is It Really?" - APPSO

"It's a Bull Market! 'Ox Coming' Runs 5 Trillion Tokens..." - AI DeepTech

"Mysterious 'Ox Coming' Large Model Goes Viral, Free for Limited Time" - Jiqizhixin

"Mysterious 'Ox Coming' Large Model Dominates Foreign Discussions, Outperforms Fable and GPT-5.6 in Some Tests" - DeepTech

"Mysterious 'Ox Coming' Shakes Up Silicon Valley! Google's Desperate Claim Gets Mocked Online" - Xinzhiyuan

"Anonymous Ox Coming Large Model Linked to Zhipu, Though Some Suspect Cursor Trained GLM Open-Source" - QbitAI

"Usage Tops Charts! Whose Family Does the Mysterious Ox Coming Large Model Belong To?" - East Money

"Mysterious New Model 'Ox Alpha' Storms OpenRouter, Outperforms Fable 5? Internet Guesses Zhipu or Xiaomi" - InfoQ

THE END

Copyright and Disclaimer

1. Content Copyright: Except for quoted public data, policies, and cases, all content is original. Professional data comes from authorized databases and government websites. Cases are compiled from real events.

2. Image Licensing: Images are either original content, officially licensed, or AI-generated. For network images with unclear copyright, rights remain with the original authors and will be removed upon request.

3. Republication Rules: Unauthorized reproduction is prohibited. Republication requires full attribution and author credit.

4. Liability Disclaimer: This article represents the author's observations and industry commentary based on public information. Content is for reference only and does not constitute professional advice. Users assume all risks from application. Xinzhidongneng reserves final interpretation rights.