Ox Alpha Unmasked as Z.ai's Open-Weight GLM-5.3-Flash Model

The stealth model gets a name

A free AI model with no listed maker kept developers guessing for about a week. Then Z.ai, the Beijing company also known as Zhipu, said it built Ox Alpha. The model is an anonymous preview of GLM-5.3-Flash, the newest model in Z.ai's GLM line, which Z.ai ran on OpenRouter and OpenCode to collect user feedback before launch.1214 Bloomberg broke the confirmation. Other outlets then repeated it, and Z.ai followed with a blog post and a public release of the model weights on Hugging Face.1112

The result surprised few people who had been tracking the story. Before Z.ai said anything, much of the developer community had already traced the model to the GLM family. TechCrunch's headline called it a surprise, but its own report said the attribution matched what many had expected.11 Still, the episode shows how new AI models are now released, tested and marketed. It also shows how much ground Chinese open-weight labs have gained against the expensive frontier models from US companies.

A week of free compute and forensic sleuthing

Ox Alpha went live on OpenRouter on August 20 under a provider simply labeled "stealth." OpenRouter stated that it only routed the traffic and did not build or operate the model.38 Its listing described a reasoning model for coding, long-running agent tasks and production work, including workflows that mix text with visual material.7 The specs stood out: a context window of about one million tokens, a 131,072-token output limit, and input in text, images and video.89 One explainer said this was the first stealth release to advertise video input.8

The free offer was just as unusual. OpenCode said the model would be free for about a week with near-unlimited usage, and that its provider could handle 100 trillion tokens a day.212 Stripe CEO Patrick Collison called it "very impressive," which brought the model a much wider audience.612

The identification work makes a strong case that technical probes now beat behavioral guessing. Asking a model about its views or its refusals can be thrown off by a system prompt. Tokenizers and error handling are much harder to disguise.5 A community fingerprinting tool ran nine infrastructure probes and found Ox Alpha matched GLM-5.3 on six of them. The same writeup said one independent tester found token counts identical to GLM-5.3 across 25 prompts, apart from a fixed 75-token offset.18 Testers also spotted an error code previously linked to Z.ai's API, and video-token usage that matched Z.ai's GLM-5V-Turbo exactly.318 One Reddit thread reported that the model introduced itself as a Z.ai GLM in its first reply.16

There were other theories. Some people suggested Xiaomi's MiMo, Tencent's Hy4, MiniMax, or even a router switching between several models.8 A few guessed Gemini, and Business Insider reported that some developers thought Microsoft might be involved.314 On Chinese social media, users named it the "Niu Lai Model," after a low-budget animated film about a calf that had become a meme.14 One analysis said prediction markets had put about 90% odds on Z.ai before the confirmation.4

What Z.ai actually shipped

According to one detailed technical writeup, GLM-5.3-Flash uses the MIT license and a mixture-of-experts design with 320 billion total parameters, 18 billion of them active.18 Business Insider reported a context window of up to one million tokens and multimodal input. OpenRouter's listing for the named model now shows 1,310,720 tokens.1215

Price may matter even more than the specs. Bloomberg reported a list price of $0.15 per million input tokens and $0.50 per million output tokens, which puts the model next to DeepSeek among low-cost, efficient options that are pulling users away from premium offerings like Anthropic's.13 Z.ai's shares rose as much as 12% after the confirmation.13 At least one third-party router listed the model at half that price as a launch promotion.18

The free preview has ended. OpenRouter's stealth page now says the free period is over, and the old preview model IDs are no longer the recommended routes.115

The chip claim may be the bigger story

The model itself is impressive. Z.ai's claim about the hardware it ran on may carry more weight. The company says all Ox Alpha traffic during the test ran on Chinese AI chips. It admitted those chips are limited in memory capacity and bandwidth, and said its engineers made up for this with software optimization and better workload distribution.14 One writeup puts the fleet at about 100,000 domestic chips serving 100 trillion tokens a day.18 The Rundown reports Z.ai's claim that this setup now serves tokens at roughly the same cost as Nvidia hardware.17

If that holds up, it matters a great deal. Washington has limited China's access to advanced AI chips.14 A frontier-class model serving global traffic on domestic hardware would suggest those controls are doing less to slow China's inference capacity than intended. One analysis links this to Nvidia CEO Jensen Huang's warning that export controls could push China to build its own stack.18 But this is a company claim, not an audited figure. One writeup noted that many outlets have repeated it without independent checks.18 Until someone outside Z.ai verifies it, the claim should be read as an important signal, not a settled fact.

Where the coverage diverges: benchmarks

The reports differ most on performance. TechCrunch and others described Ox Alpha as climbing benchmarks and leaderboards against the best models.1119 The Rundown said it took OpenRouter's top spot with roughly double the usage of second-place DeepSeek, which the platform called its biggest debut ever.17 The Rundown also cited a score of 57 on the Artificial Analysis Intelligence Index, at about a tenth of the per-task cost of similarly ranked models.17

Other reports are more cautious. A detailed tracker said that, at the time it wrote, Ox Alpha had no entry on independent leaderboards. It called the published figures, including the Index score of 57 and coding results such as DeepSWE, vendor-reported.18 Another analysis found that early test results were inconsistent. Some quick tests had the model beating rivals. Retests on private benchmarks built to avoid training-data contamination put it closer to the GPT 5.5 to 5.6 range, slightly below GPT 5.6 Sol-mid.4 One explainer warned that the most-shared early benchmark came from a very small test run by a user.9 Bindu Reddy's hands-on notes praised the free access, long context and multimodality, but said the model "spins a lot."3 TPS Report noted that neither GLM-5.3's claimed parity with Anthropic's Fable 5 nor Ox Alpha's leaderboard position had been checked with standardized, reproducible tests.20

My reading is that "tops the leaderboards" mostly describes usage, not intelligence. Huge OpenRouter traffic shows how much developers wanted a capable model that cost nothing. On quality, the evidence points to a very good, cheap, frontier-adjacent model, not one that beats the best closed systems. Early community scores on one benchmark put Ox Alpha just below GLM-5.3, which fits that view.18 What stands out is how much capability Z.ai offers for the price, more than the capability on its own.

The stealth launch is now a standard playbook

Ox Alpha follows a clear pattern. Z.ai previously tested GLM-5 anonymously as "Pony Alpha" and claimed it after about five days.1418 According to one tally, all four anonymous releases before Ox Alpha were later traced to Chinese labs: Zhipu's GLM-5, Xiaomi's MiMo-V2-Pro, Ant Group's Ling-2.6-flash and Meituan's LongCat-2.0.18 The appeal is easy to see. An unbranded model gets judged only on its outputs, which helps it avoid skepticism toward Chinese models, and the reveal then creates a second wave of attention. One analysis said Z.ai credited the Pony Alpha approach with showing that Chinese models could compete at the frontier.4

The approach has costs. During the anonymous period, OpenRouter's listing said prompts and completions were kept by a provider no one could name, though not used for training.2 OpenCode's announcement promoted zero data retention, so the policy may have depended on which route a developer used.8 The Next Web argued that free access was really paid for with data, and advised users not to send anything sensitive.2 Now that the maker is known, teams that pasted proprietary code into Ox Alpha during the preview know where it went. They had no way of knowing at the time.

Why it matters

The Ox Alpha reveal brings three trends in AI together. First, Chinese labs keep releasing open-weight models that come close enough to the frontier to threaten the pricing of OpenAI and Anthropic.11 Second, Z.ai is presenting software optimization on domestic chips as a way around US export controls.14 Third, anonymous previews have become a normal launch tool for building reputation through raw performance before a brand is attached.

The most practical change is for developers. A model that one analysis rates near GPT 5.6 Sol-mid can now be downloaded and run locally.18 That turns ongoing per-token costs into a one-time hardware cost for anyone willing to host it.18 Developers still need to check the license, the infrastructure requirements and serving quality before moving to self-hosting.15