Ox Alpha Revealed: Z.AI Releases It Open-Weight as GLM-5.3-Flash

On the evening of Wednesday, August 26, the mystery ended the way the forensics said it would. Z.AI — the Beijing lab behind the GLM series — released the model it had been testing under a mask as an open-weight product, under the name GLM-5.3-Flash, exactly the sub-name tokenizer forensics and prediction markets had converged on. Ox Alpha, the anonymous model that has topped usage charts since it surfaced on August 20, is now a published, self-hostable model: MIT license, 320B total parameters with 18B active per token, the 1,048,576-token context window and text/image/video input it advertised on the listing, and weights on Hugging Face. The reveal also answered the biggest puzzle of the anonymous week — how a model nobody owned could move 100 trillion tokens a day: Z.AI says the serving ran entirely on roughly 100,000 domestic Chinese AI chips, the first time Chinese silicon has carried frontier-scale global traffic. The capacity had also spawned the week's most popular wrong guess — at that scale, many observers, encouraged by cryptic posts from Google DeepMind researchers, argued Ox Alpha had to be Gemini running on NVIDIA hardware; the reveal corrected that too. The same night, Alibaba shipped Qwen3.8-Flash-Next, an open-weight preview of its Qwen4 architecture — one evening, two frontier-class open-weight releases from Chinese labs. The four anonymous releases before Ox Alpha were all eventually claimed by Chinese labs: Zhipu AI's GLM-5, Xiaomi's MiMo-V2-Pro, Ant Group's Lingxi Ling-2.6-flash, and Meituan's LongCat-2.0. Ox Alpha is no longer a platform-exclusive experiment; it is a model developers can self-host.

Every figure in this piece is the operator's own claim, data from the platform's own listing, a public announcement, or community fingerprinting. The DeepSWE numbers are independent researcher Ben Davis's community measurements — a full 113-task run, not an official leaderboard entry, though the ~63% figure now lines up closely with Z.AI's own vendor-reported DeepSWE v1.1 score of 63.4. The identity question is closed: on August 26 Z.AI released Ox Alpha as an open-weight model under the name GLM-5.3-Flash — MIT license, 320B total parameters with 18B active — confirming the sub-name that tokenizer and video-encoder forensics had converged on. Architecture, parameter count, and pricing are now company-published; what remains vendor-stated rather than independently verified is the benchmark suite and Z.AI's claim that the anonymous week's serving ran on roughly 100,000 domestic Chinese AI chips at a per-token cost comparable to mainstream NVIDIA hardware — a company claim that multiple outlets have now repeated, not an audited figure. The early Gemini hypothesis — which the 100T-token/day capacity had spawned and cryptic posts from Google DeepMind researchers encouraged — was wrong, and the release settled it. The quality rating attributed to analyst teortaxesTex — and his suggestion that a further-trained Ox Alpha could be the workhorse for a stronger GLM 5.5 — is that analyst's own assessment from a social post, not an independent measurement or a Zhipu statement. The looping-animation demo described below is likewise a community report — a developer's post on X, echoed on Chinese developer forums — not a vendor capability claim, and the operator has announced no such feature.

What we know — and what we don't

The fast version:

• The operator describes Ox Alpha as "a frontier model built for efficient coding, sustained agentic work, and real-world production use" — a reasoning model aimed at long-horizon software engineering and workflows that mix text with visual context.

• It has a 1,048,576-token (1M) context window, a 131,072-token output limit, and takes text, image, and video input — the first of the anonymous releases to advertise video input, and a capability the GLM-5.3-Flash release confirms as native rather than bolted on.

• It was free for one week: $0 per million tokens in and out, with the operator claiming "generous rate limits, near-unlimited usage" and 100 trillion tokens per day of serving capacity — capacity the reveal would later say was provisioned on roughly 100,000 domestic Chinese chips.

• Confirmed: on August 26 Z.AI released Ox Alpha as an open-weight model under the name GLM-5.3-Flash — MIT license, 320B total parameters with 18B active — the exact sub-name tokenizer, video-encoder, and error-code forensics plus prediction markets had converged on. Independent analyst teortaxesTex rates it at roughly GLM-5.3's level, vision aside, and suggests a further-trained Ox Alpha could be the workhorse behind a much stronger GLM 5.5.

• The flash-multimodal reading now has a demo behind it: asked to generate a looping animation of a pelican riding a bike and opening a phone that plays the same animation, Ox Alpha answered with a self-contained looping animation as a single HTML/SVG file — a community demo, not a vendor claim.

• Early activity data on the platform already shows real production traffic, including a coding agent from Nous Research and the Zed editor.

• The company has now shipped it — on August 26 Z.AI released Ox Alpha as the open-weight GLM-5.3-Flash under the MIT license, ending the identity, spec, and price questions at once: 320B total parameters with 18B active, natively multimodal, with a Z.AI list price of $0.15 in / $0.50 out per million tokens and a 50% launch promo stated to run only through September 9 — a date now eight days past, with the discounted rate still the one being served. Z.AI also disclosed that the anonymous week's 100T-token/day serving ran on roughly 100,000 domestic Chinese chips, a first at that scale. What the reveal did not include is an independent benchmark — the model's scorecard is vendor-reported — so the most representative independent number remains Ben Davis's full 113-task DeepSWE run at roughly 63%, on par with GPT-5.6 Sol mid.

What Ox Alpha is

The platform's listing describes Ox Alpha as "a reasoning model designed for coding, sustained agentic work, and production workloads — long-horizon software engineering, complex reasoning, and workflows that combine text with visual context." That is a specific positioning: it is built for long-running agent loops and for code, not for general chat. The 1M context is the give-away — that is enough room for a multi-hour agent session or a large repository — and the 131K output cap allows for long single generations. It supports tool and function calling and structured JSON output, which is the rest of the agentic checklist.

The video input is the most distinctive item on the sheet. None of the earlier anonymous models advertised it, and a model that can take video frames as well as text and images is aimed at a different class of workload than the chat-tuned releases that preceded it. It is also, as the forensics would later show, one of the strongest identity tells a model can carry. All of this — the description, the specs, the speed figures — came from the operator and the platform's listing. The GLM-5.3-Flash release confirmed the headline specs (320B-A18B, native multimodal); the speed and capacity figures remain vendor claims.

The multimodal story gained a concrete demo this week. Developer Chetaslua — whose server-side probes are part of the fingerprinting trail — posted a test in which he asked Ox Alpha to make a loop of a pelican riding a bike and opening a phone that plays the same animation back: a self-referential loop inside the loop. His report has the model delivering a single-file HTML/SVG animation — a pelican with two-segment legs that genuinely pedal, wheel speed synced to ground speed, a parallax background, and the phone-in-the-loop payoff. Chinese developer forums separately ran their own pelican-bike prompts and discussed the result, so the demo is not a single unverifiable claim. Because the output is code, it is consistent with the platform's text-output-only spec: multimodal understanding and creative code generation working together, not native video synthesis. Chetaslua's takeaway — that this is the payoff of multimodality, and that a capable open-source model will arrive with it — is his own forecast, not a vendor statement; the operator has announced nothing.

The free-for-a-week deal

The promotion was aggressive. Free for the week it ran, "generous rate limits, near-unlimited usage," and a claimed 100 trillion tokens per day of capacity, with the operator's note inviting users to "see what you can do." Taken literally, 100T tokens a day would put this operator among the largest inference providers anywhere — the claim reads less like a spec and more like a stress-test challenge, which fits the pattern: anonymous releases are how a lab gets frontier-scale real-world traffic without putting its name on the door. The confirmed reading made the scale claim concrete rather than just easier to square: Z.AI disclosed that the 100T-token/day serving ran on roughly 100,000 domestic Chinese AI chips — the first time a frontier-class release has been served at that scale without NVIDIA hardware. That disclosure is still a company claim, now repeated by multiple outlets but not independently audited, and it carries a second, softer vendor assertion: Z.AI says per-token cost on the domestic cluster is comparable to mainstream NVIDIA GPUs.

The other side of "free for a week" was that the price that followed it was unknown — the reveal answered that. Z.AI's published list price for GLM-5.3-Flash is $0.15 per million input tokens, $0.50 per million output tokens, and $0.03 per million cached input. A 50%-off launch promo was stated to run through September 9, 2026, cutting that to $0.075 / $0.25. Eight days past that date, the discounted rate is still the rate being served: re-read on September 17, 2026, the z-ai/glm-5.3-flash model page prices input at $0.075, output at $0.25, and cache reads at $0.0173 per million tokens, with no promotional label on it. Against GLM-5.3's $1.40 / $4.40 list, that is roughly a twentieth. The practical consequence is that the September 9 end date is stale rather than binding — a developer budgeting on $0.15 / $0.50 is budgeting on a figure nobody is currently being charged. What the pricing pages do not settle is why. Z.AI's own model list still carries $0.15 / $0.50 for GLM-5.3-Flash, so whether the promo was extended, made permanent, or simply left running is not something we can source; the discounted rate is the observed fact on this date, and the cause stays open.

The first full benchmark — and it lands on par with GPT-5.6 Sol mid

Ox Alpha still has no entry on independent trackers such as Artificial Analysis or LMArena — the word "frontier" is the operator's own, and the benchmark figures Z.AI published with the GLM-5.3-Flash release (DeepSWE v1.1 63.4, AutomationBench 48.8, an Artificial Analysis Intelligence Index of 57) are vendor-reported, not independent scores. What has changed since launch is that community-measured numbers are starting to circulate — and the picture has narrowed. On Kingbench, early runs put Ox Alpha at 87.5, just below GLM-5.3's 91.25. On DeepSWE, independent researcher Ben Davis first ran a 10-task subset and reported 80% Pass@1, ahead of Claude Fable 5 (65%), GLM-5.3 and Grok 4.6 (62%), and GPT-5.6 Sol (52%). He has since completed a full run against the entire 113-task DeepSWE set, and the corrected result is roughly 63% — more or less on par with GPT-5.6 Sol mid. The 80% subset was the eye-catching number, but ten tasks is under 9% of the benchmark; Davis himself flagged the full-set figure as the one that "makes way more sense." These are community-measured figures, not a vendor benchmark and not an official leaderboard score — but the full run is the most representative number anyone has gotten out of the model, and it now lines up closely with Z.AI's own vendor-reported DeepSWE v1.1 score of 63.4 — a useful cross-check, even though the company's number is a claim rather than a measurement.

One comparison matters more now than it did at launch. DeepSeek V4 Pro 0813 — the GA build of DeepSeek's flagship that shipped on August 13 — reports a DeepSWE of 62.7 on DeepSeek's own benchmark card, against 12.8 for the earlier DeepSeek V4 Pro Preview. Both figures are vendor-reported, unreproduced by an independent lab as of this writing. Read alongside GLM-5.3-Flash's ~63% community full-set run and Z.AI's own DeepSWE v1.1 of 63.4, DeepSeek's 62.7 puts the two best-known Chinese coding-agent claims in the same low-60s band. The ten-point gap that looked like Ox Alpha's calling card was measured against DeepSeek V4 Flash 0731, the cheaper small model; against DeepSeek's flagship, GLM-5.3-Flash lands level, not ten points clear.

The other lesson is about the number itself. Analyst teortaxesTex — who has tracked these estimates since the anonymous week — notes that the first report about Ox Alpha gave 80% DeepSWE too: that was Davis's eye-catching 10-task subset, and the end result on the full 113-task set was 63%. With DeepSeek V4 Pro 0813's official 62.7 in the same band, his observation is that nothing has come close to a legitimately reproduced 80% yet — the 80% figure keeps appearing early in a model's public life and keeps settling in the low 60s once the full set is run. That is his analyst reading, not a measurement, but it is the right lens for the next DeepSWE headline, including any about GLM-5.3-Flash itself.

What also exists is usage. The platform's activity data shows real production traffic within the first hours — including Hermes Agent, the agentic coding project from Nous Research, and the Zed editor. Both are exactly the "sustained agentic work and coding" use cases the model is positioned for. It is not a score, but it is a signal: developers with real workloads decided a free, frontier-class, anonymous gamble was worth wiring up. Before the full DeepSWE run landed, that early adoptership was the only evidence there was; the ~63% full-set figure now gives the model a benchmark anchor to weigh against it.

The data-retention fine print

The free-week announcement touts "zero data retention." The model's listing on the platform tells a more qualified story: prompts and completions are retained by the provider but not used for training, under the platform's stealth-model terms. The difference matters if you are about to paste proprietary code into a 1M-token window owned by a provider that was anonymous until Z.AI stepped forward on August 26. Treat "zero data retention" as marketing until Z.AI publishes terms that say so in so many words — and route anything you would not want logged through a separate, attributable model.

The anonymous-model playbook

Ox Alpha is the fifth anonymous release in six months, and the previous four resolved the same way — an anonymous debut, a burst of free traffic, then a company stepping forward:

• Pony Alpha (February 2026) — confirmed by Zhipu AI about five days after launch as GLM-5, its 744B-parameter MoE flagship.

• Hunter Alpha (March 11) — a free, trillion-parameter model with a 1M context that became the speculation story of the spring, widely guessed to be Deep​Seek V4; revealed as Xiaomi's MiMo-V2-Pro, with the companion Healer Alpha identified as MiMo-V2-Omni.

• Elephant Alpha (April) — a free, efficiency-focused text model that Ant Group claimed as Lingxi Ling-2.6-flash about two weeks after it appeared.

• Owl Alpha (late April) — an agent-focused model with native tool calling and a ~1M context that Meituan confirmed as LongCat-2.0 on June 30, the first trillion-parameter model trained and served entirely on domestic Chinese chips.

• Ox Alpha (August 20) — revealed August 26 as GLM-5.3-Flash, an open-weight, MIT-licensed 320B-A18B model; the identity, license, and specs were all official on the same day the mask came off.

Why launch a good model behind a mask? The standard reading, which the Hunter Alpha cycle made concrete, is that anonymity removes brand bias from evaluations, and the free usage crowdsources real-world evaluation data at frontier scale while everyone argues about the owner. As one analysis of the pattern put it, "the lack of a brand is the marketing." The extra win is speed: an anonymous model can reach a global developer audience in hours without a launch event, and the mystery itself becomes the distribution.

Who is Ox Alpha?

Until the August 26 release, no company had claimed it and Zhipu AI had said nothing — but the hard-evidence situation changed within 48 hours of the model's debut, and the forensics below are why the reveal — as GLM-5.3-Flash — landed with so little surprise. The capacity figure briefly worked against the forensics: 100 trillion tokens a day looked like the signature of a top-tier lab on NVIDIA silicon, and the theory that Ox Alpha was a new Gemini — encouraged by cryptic posts from Google DeepMind researchers — had real momentum until the tokenizer evidence, and then Z.AI's own release, closed it. The strongest signal is the tokenizer. A community-built fingerprinting tool, modelprint, ran nine infrastructure probes against Ox Alpha and found it matched Z.ai's GLM-5.3 on six, with all four normalized tokenizer counts matching the GLM family while the best non-GLM candidate managed two of four. Independent researcher Ben Davis reported Ox Alpha's token counts matched GLM-5.3 exactly across 25 test prompts, with only a constant +75-token difference he attributed to a hidden wrapper. Tokenizer matching is the hardest identity signal to fake — a model cannot change how it segments text without retraining.

The video path is the second signature. Ox Alpha's video token consumption matched GLM-5V-Turbo exactly across four controlled samples — the same frame-sampling, duration-scaling, and resolution mechanics — while MiMo v2.5, Qwen 3.8 Max, and GLM-4.6V all diverged. That is a strong tell: video handling is deep in the model, not a bolt-on. The rest of the fingerprint stack lines up with the same reading: Ox Alpha rejects audio input the way GLM-5V does, shares GLM's characteristic "1301" error code, emits a similar output style (about 1.3 emojis per 1,000 characters), carries the same restricted parameter and API configuration that appeared on the platform's GLM-5.3 listing two days before Ox Alpha launched, and has a knowledge cutoff near November 2025.

The identification has a precedent in Zhipu's own playbook: Pony Alpha, the anonymous February release, turned out to be GLM-5, claimed about five days after launch. The analyst reading circulating this week — that Ox Alpha is GLM-5.3, presumably the Flash model — was echoed by Pliny the Liberator, who said his agent had confirmed "Ox-alpha is from Zai, GLM-5.X family." Independent analyst teortaxesTex makes the same call on quality and adds a timing claim: he rates Ox Alpha at roughly GLM-5.3's level, vision aside, and finds the turnaround the most striking part — compressing the text-only GLM-5.3 and shipping it with working vision within days of the August 14 launch. That is one analyst's assessment, not an independent score, and he argues it should give pause to the "China releases models hot off the press" narrative. His newest post layers a roadmap guess on top of that read: the GLM-5 update cycle may be short; Ox Alpha is close enough to GLM-5.3 that using it as a subagent barely makes sense — "just run ox @4 instead" is his shorthand for running the model directly; and a further-trained Ox Alpha would make a lot of sense as the workhorse, in his word a "beast of burden," for a much stronger GLM 5.5. That is one analyst's speculation about a roadmap, not a Zhipu statement. The sub-name question, by contrast, is settled: on August 26 Z.AI released Ox Alpha as GLM-5.3-Flash. The wrinkle that once kept the Flash label on the "presumed" side — GLM-5.3 launched August 14 as a text-only model, while Ox Alpha advertises video input — is exactly what the release resolved: GLM-5.3-Flash is a distinct, natively multimodal model rather than the same text-only one. Alternative theories — Xiaomi's MiMo team, DeepSeek — have been floated and largely discounted on the tokenizer and video evidence, and the release settles the family question outright. Davis's argument for caring about the Flash label turned out to be the practical one: because Ox Alpha really is GLM-5.3-Flash performing at GPT-5.6 Sol mid level, it can now run locally — the weights are on Hugging Face and SGLang, vLLM, and TokenSpeed serve it — which turns a mid-tier frontier coding model into a one-time hardware cost instead of a per-token bill for anyone willing to host it.

The geopolitical framing is the analysts', not the evidence's: "This puts even more immense pressure on US Frontier Labs, and the gap with China is shrinking." That is an interpretation of what a free Zhipu-tier model running at frontier scale means for the competitive order — and the hardware disclosure gives it an edge the analysts did not have. Serving 100 trillion tokens a day on roughly 100,000 domestic chips is the first large-scale demonstration that a Chinese stack without NVIDIA silicon can carry frontier inference traffic — the scenario NVIDIA CEO Jensen Huang warned about on the Dwarkesh Podcast this spring, when he argued that export controls would accelerate China's NVIDIA-independent stack and that a frontier model running best on domestic Chinese hardware would be a "horrible outcome" for the United States. Z.AI's cost-parity figure is still a vendor claim, but the event itself is real. The same evening made the point concrete on the model side too: as Z.AI released GLM-5.3-Flash, Alibaba shipped Qwen3.8-Flash-Next, its open-weight preview of the Qwen4 architecture. Two Chinese frontier-class open-weight releases in a single night is the concrete shape of what observers called "a big day for Chinese open source."

Should you build on it this week?

The free week is over, but the test is still cheap: the rate actually served on September 17 is $0.075 in / $0.25 out per million tokens, not the $0.15 / $0.50 list price — the 50% launch promo that was stated to end September 9 is still in force eight days later. If you do long-horizon agentic work, code on long contexts, or run video-heavy pipelines, that is exactly the intersection where Ox Alpha is positioned — and with the weights open, you can run the test on your own hardware rather than at an anonymous platform's sufferance. Two cautions. First, the "frontier" label is still a claim: GLM-5.3-Flash's scorecard is vendor-reported, and the one solid independent number — Davis's ~63% DeepSWE run — is strong but a long way from the viral 80% of the early 10-task subset. Second, the $0 price was time-boxed, and the reveal priced what comes after — cheap for the capability, but no longer free. What the open-weight release removes is the dependence: the files are out, so the model can be self-hosted rather than rented. That is the shape of a test, not a dependency.

The production-safe way to run a test like that is a fallback chain: the experimental model answers when it is healthy, and the request rolls to a proven model the moment it stalls, errors, or disappears. That is what a routing layer is for. The reveal makes the fallback choice trivial: Ox Alpha is GLM-5.3-Flash, and the model itself is live on OrcaRouter under its real name at $0.075 in / $0.25 out per million tokens, with cache reads at $0.0173 — the 50%-off launch rate that was stated to end September 9 and, as of September 17, is still the price on the page rather than the $0.15 / $0.50 list figure. It is passed through with 0% markup, one API across 200+ models, with automatic failover so a launch-week traffic spike does not become your outage. Audition the new model in front with a proven fallback behind it in the chain; when a price changes, the number you see on OrcaRouter is the number you pay, the same day. Or skip the API entirely — the weights are open, so self-hosting GLM-5.3-Flash behind the same chain is on the table too.

What to watch

Six things will tell the rest of the story. The independent benchmark: GLM-5.3-Flash's scorecard is vendor-reported, Davis's ~63% DeepSWE run is the only solid independent number so far, DeepSeek V4 Pro 0813's official 62.7 now sits in the same band, and an Artificial Analysis or LMArena entry — or an independent head-to-head — would be the final check on whether "frontier" is a fact or a tagline. The price trajectory: the steady-state question has an answer on the evidence so far — the 50% promo was stated to end September 9, yet on September 17 the discounted $0.075 / $0.25 rate is still what is served, so the promo figure rather than the $0.15 / $0.50 list price is the number to budget on; what stays unresolved is the cause, since Z.AI's own list price has not moved. The hardware claim: Z.AI says the anonymous week ran on roughly 100,000 domestic chips at cost parity with NVIDIA — the first public test of that at scale is whether the claim survives independent scrutiny, and whether the domestic stack becomes the default for the GLM line. The adoption math: whether the platform-topping traffic Ox Alpha drew under the mask converts into self-hosted and API deployments now that it has a name, a license, and a price. The sequel: whether GLM-5.3-Flash casts Ox Alpha as a stepping stone — teortaxesTex's hint is that a further-trained version becomes the workhorse for a much stronger GLM 5.5, which would make this release the foundation of the next GLM. And the reaction: with Qwen3.8-Flash-Next landing the same night and a domestic-chip reveal behind it, the pressure is on US frontier labs — and on the export-control debate — to answer; watch whether a fast-follow, a price move, or a policy response lands on the other side of the gap.

The honest verdict: Ox Alpha was an unusually cheap experiment in both directions. The platform tested whether an anonymous frontier-class model at $0 could win real production traffic in a week — it did, convincingly enough that Z.AI not only stepped forward on the sixth day but shipped the weights as an open model the same night. You get to test whether the model deserves a place in your stack, now without the anonymity risk: GLM-5.3-Flash is a named, MIT-licensed, self-hostable model at a published price, and the hardware question got an answer too — Z.AI says the 100T-token/day serving ran on roughly 100,000 domestic Chinese chips, company-stated but now widely reported. What remains to be learned is whether "frontier" survives independent measurement, whether Z.AI's domestic-chip cost-parity claim holds up at scale, why the 50% launch promo is still the price being served eight days past its stated September 9 end date, and whether the reveal casts GLM-5.3-Flash as the workhorse for a stronger GLM 5.5, as teortaxesTex hints. Run the experiment, but run it with a fallback underneath.