Ox Alpha wasn’t just Chinese — it was fully powered by Chinese chips.
Z.ai founder and CEO Jie Tang put that fact front and centre in a post celebrating the model’s run on OpenRouter, where the stealth release that had the AI community guessing for the better part of a week turned out to be the company’s own GLM-5.3-Flash. “Ox Alpha = GLM-5.3 Flash AA = 57, 1/100 frontier price, Powered by pure Chinese chips. Delivered nearly 20% weekly token share (no. 1) on OpenRouter. Thanks to all for the support,” Tang wrote, tying together three separate claims that each carry weight on their own: a benchmark score in the same range as frontier Western models, a price that undercuts them by two orders of magnitude, and an infrastructure story built entirely on domestic silicon.
The 57 that Tang references is Ox Alpha’s score on the Artificial Analysis Intelligence Index, a composite benchmark that has become one of the more widely cited yardsticks for comparing model capability across labs. Landing in that territory puts a model that costs a fraction of what OpenAI or Anthropic charge within shouting distance of systems that are priced nowhere close to it, which is precisely the pitch Z.ai has been building toward with its GLM 5 line for months now.
From Mystery Model To Named Release #
Ox Alpha first showed up anonymously on OpenRouter and OpenCode on August 20, offering a 1-million-token context window, text, image and video input, and a claimed serving capacity of 100 trillion tokens a day, all completely free for roughly a week. The scale of the giveaway is what triggered the guessing game in the first place. A model good enough to compete with frontier systems, handed out for nothing at a volume that size, is not something most labs can casually afford, and the community spent days running tokenizer fingerprinting tests and probing the model’s behaviour to figure out who was footing the bill. As OfficeChai reported at the time, the tokenizer evidence pointed toward Z.ai’s GLM family well before the company confirmed anything, with Ox Alpha lining up against GLM’s tokenizer on essentially every probe researchers threw at it.
Z.ai ended the guessing this week, confirming that Ox Alpha is GLM-5.3-Flash, a 320-billion-parameter hybrid model that only activates 18 billion parameters per token. That sparse activation is the mechanical reason the model can be priced so aggressively in the first place, and it is reported to beat GLM-5.2 across the board while landing at roughly a tenth of that model’s own price, which was already cheap by American standards.
The Chip Story Is The Real Headline #
What makes this release different from a routine model drop is the infrastructure claim sitting underneath it. Z.ai says GLM-5.3-Flash was served over its entire free week on a cluster of domestically produced Chinese AI chips, running on a custom inference engine the company built on SGLang. The company has said it doesn’t intend to disclose exactly which chips were used, only that the individual units are limited in compute and memory relative to what Nvidia ships, and that GLM-5.3 itself was used to help optimise the serving stack for its own smaller sibling, a feedback loop the company frames as central to hitting the throughput numbers it needed.
That is a notable claim to stand behind in public, because serving a genuinely popular model at 100 trillion tokens a day is an operational problem, not just a training one. Training a capable model on domestic hardware and proving that a single lab can absorb the kind of viral traffic Ox Alpha pulled in, purely on that hardware, are two different achievements. Z.ai has form here. Its GLM-5.1 and GLM-5.2 releases were both trained entirely on Huawei’s Ascend chips, with the company sitting on the US Entity List since early 2025 and effectively barred from buying American accelerators. What Ox Alpha adds to that record is the inference side of the equation, at a scale large enough that it was impossible to fake or quietly outsource.
The Compute Question Everyone Got Wrong #
For most of the week Ox Alpha was live, one theory kept surfacing in developer forums and on X: that whoever was behind it must be burning through borrowed or rented Nvidia compute, because no Chinese lab could plausibly sustain that kind of free, high-volume traffic on domestic chips alone. DeepSeek was floated and then largely ruled out on exactly this basis, with observers pointing to the company’s decision to raise prices on its V4-Flash model as evidence it didn’t have the spare capacity to give away tokens at that rate. The assumption baked into that entire line of speculation was that China’s chip stack simply wasn’t there yet, and that any model performing this well, this cheaply, for this many people, had to be leaning on Western silicon somewhere in the pipeline. Z.ai’s confirmation puts an end to that narrative. A lab served a model with frontier-adjacent benchmark scores, at a volume large enough to top OpenRouter’s weekly rankings, entirely on chips built inside China, for a full week, without the wheels coming off. That is a different claim than training a model on Ascend hardware in a controlled environment — it is a claim about handling real, unpredictable, high-volume public traffic on hardware that the company itself admits is more constrained than Nvidia’s, and doing it long enough for outside observers to actually notice and rank it.
What The OpenRouter Numbers Actually Show #
The token-share data backs up Tang’s assertions. Z.ai’s slice of weekly tokens on OpenRouter had been bouncing between 3% and 9% for most of the summer, well behind DeepSeek, which frequently sat in the mid-teens to low-twenties. That changed sharply once Ox Alpha entered the mix under its stealth branding: Z.ai’s combined share jumped to 24% for the week of August 17, then to 23% the following week, with GLM-5.3-Flash alone accounting for 19 percentage points of that once the model was unmasked. DeepSeek, by comparison, slipped to 16% over the same week. For the first time in the period OpenRouter tracked, a Z.ai model pushed the company’s overall share past DeepSeek’s, and it did so while the company was giving the model away for free.
Why This Matters Beyond One Ranking #
None of this happens in a vacuum for OpenAI and Anthropic, both of which are working through confidential IPO filings while trying to convince public-market investors that their pricing power will hold. Chinese labs charging a small fraction of what US labs charge for comparable coding and reasoning work has already become a talking point on Wall Street, and Ox Alpha adds a specific new wrinkle to it: this isn’t just a price gap anymore, it’s a demonstration that the underlying compute gap investors have been pricing in may be narrower than assumed. If a lab can serve frontier-adjacent performance on domestically produced chips at a scale big enough to top a public leaderboard, the argument that Chinese labs are compute-constrained and therefore permanently a tier behind on distribution gets harder to make with a straight face.
That argument has already been showing up in analyst notes and investor commentary ahead of the IPOs. Big Short investor Steve Eisman has said publicly that cheap Chinese pricing on models like Kimi K3 would leave him “petrified” if he ran either lab, warning of a price war just as both companies aim for valuations near or above $1 trillion. Z.ai’s own GLM-5.2 had already climbed ahead of every Google model on the Artificial Analysis Intelligence Index while training entirely on Ascend chips, at an estimated cost of around $25 million. Ox Alpha’s free week, and the chip claim attached to it, is another data point in the same direction, and one that arrives at a particularly inconvenient moment for two labs trying to convince public investors that their technological lead is worth a premium few enterprise buyers now seem willing to pay without a fight.
Tang’s post reads like a victory lap, and on the numbers he’s entitled to one. But the more consequential part of his message isn’t the ranking. It’s the quiet insistence, three times over in two sentences, that none of it required a single foreign chip.