Z.ai is the lab behind Ox Alpha, and after six days of speculation that answer landed with the dull thud of a thing everybody suspected and nobody could prove. The stealth model that appeared on OpenRouter without an owner, ate a fifth of a coding platform’s traffic and knocked DeepSeek off a leaderboard it had held for eight weeks was a Chinese GLM checkpoint the whole time.
The confirmation came on 26 August 2026, first through Bloomberg and then in a TechCrunch report by Rebecca Bellan, published at 7:19 AM Pacific. Ox Alpha, the piece confirmed, was the newest iteration of the GLM series. Within hours the weights were on Hugging Face under an MIT licence, and the model had a real name: GLM-5.3-Flash.
That sequence is worth slowing down over, because it is the third time this year a Chinese lab has run the same play and the second time it has worked at frontier scale. We covered the stealth model itself while it was still anonymous in our earlier piece on the mystery of Ox Alpha, and the fuller catalogue of releases sits in our AI models hub. This article is the sequel: who Z.ai is, what the model actually turned out to be, which community claims survived contact with the real specification sheet, and what a technology buyer should do about any of it.
Some of what follows flatters the model. A good deal of it does not. The most widely shared benchmark number about Ox Alpha was wrong by seventeen points, and the person who published it corrected it himself before the reveal — which tells you more about how AI models get evaluated in public than the score ever did.
Table of contents
- What Z.ai Confirmed and Exactly When
- Z.ai’s Ox Alpha Was GLM-5.3-Flash All Along
- Inside Z.ai’s GLM-5.3-Flash Architecture
- The Benchmarks Z.ai Actually Published
- The 80 Per Cent Claim Z.ai Never Made
- How the Z.ai Stealth Launch Actually Performed
- How Developers Unmasked Z.ai Before the Announcement
- Why Z.ai Ran the Whole Thing Anonymously
- What MIT-Licensed Weights From Z.ai Actually Change
- Z.ai Pricing Against the Frontier Labs
- Who Z.ai Actually Is
- Risks and Open Questions About Z.ai and This Release
- What Z.ai Means for Technology Buyers
- Frequently Asked Questions About Z.ai and Ox Alpha
- References and Further Reading
What Z.ai Confirmed and Exactly When
The reveal was not a launch event. It was a confirmation given to a wire service, followed by a repository going public a few hours later.
The Bloomberg confirmation
Bloomberg reported on 26 August 2026 that Z.ai had confirmed authorship of the stealth model. TechCrunch picked it up the same morning, noting that the nerds had spent the weekend buzzing with speculation over which lab was behind a model launched onto OpenRouter anonymously and already topping benchmarks and leaderboards against the best models available. The answer, Bellan wrote, was the one many had expected.
The description Z.ai gave
The characterisation Z.ai supplied is narrow and worth quoting because it sets the evaluation frame. The model is “a reasoning model designed for coding, sustained agentic work, and production workloads. It is suited for long-horizon software engineering, complex reasoning, and workflows that combine text with visual context.” Note what is absent: no general-assistant claim, no chat positioning, no consumer story. This is a model aimed squarely at engineering teams.
The weights promise, kept the same day
Z.ai said the weights would be published that evening, and they were. The repository zai-org/GLM-5.3-Flash went live on Hugging Face under the MIT licence, which is about as permissive as open weights get — commercial use, modification and redistribution, with no field-of-use restriction and no revenue threshold.
The Hugging Face detail nobody expanded on
TechCrunch added one line in passing that deserved a paragraph of its own: the GLM series is what Hugging Face recently used to defend itself against an attack from OpenAI’s own agents. The report gave no further detail, and no elaboration has followed. It is the sort of aside that will either become a case study or quietly disappear.
Why the timing was deliberate
The free preview window was due to close around 27 August. Confirming on the 26th meant the reveal landed while usage was still peaking and every developer who had spent a week building against a nameless endpoint suddenly had a name, a licence and a price list. That is a launch sequence, not an accident.
The reveal timeline
| Date | Event | Source of record |
|---|---|---|
| 20 August 2026 | The stealth/ox-alpha listing appears on OpenRouter and OpenCode, free | Platform listings |
| 22-24 August | Tokenizer and error-code forensics point at the GLM family | Community testing |
| 24 August | Viral 80% DeepSWE claim circulates from a 10-task subset | Ben Davis |
| 26 August, morning | Nebius executive names the model, then the post disappears | Roman Chernin |
| 26 August, 7:19 AM PDT | Confirmation that Z.ai built Ox Alpha | Bloomberg, TechCrunch |
| 26 August, evening | MIT-licensed weights published as GLM-5.3-Flash | Hugging Face |
Z.ai's Ox Alpha Was GLM-5.3-Flash All Along
The community got the family right and the specification badly wrong, which is a useful thing to know about crowd-sourced model forensics.
The sub-name the crowd guessed
By 24 August the consensus on X and in developer forums had converged on “GLM-5.3 Flash” as the unofficial name, derived from tokenizer matching and video-encoder behaviour rather than from anything Z.ai said. That guess turned out to be exactly right, down to the hyphen. Several competing theories — Microsoft’s unreleased MAI line, an IBM Granite checkpoint, a Google experiment, Xiaomi’s MiMo family, MiniMax — did not survive.
The Nebius slip
The single most direct piece of evidence arrived by accident. Roman Chernin, co-founder and chief business officer at Nebius, posted on 26 August that Ox Alpha “is GLM-5.3-Flash” and that it was coming to the Nebius Token Factory. The post became unavailable shortly afterwards. Infrastructure partners get told before the press does, and occasionally one of them forgets which embargo they are under.
What the Z.ai model card settled
The published card resolved the questions forensics could not reach. GLM-5.3-Flash is a mixture-of-experts model with 320 billion total parameters and 18 billion active per token, trained on a 30-trillion-token multimodal corpus. Community estimates during the stealth window had put it near 744 billion parameters with roughly 40 billion active — an overestimate of more than double on both counts.
Where the crowd was wrong
That miss matters. The whole appeal of GLM-5.3-Flash is that a 320-billion-parameter sparse model with 18 billion active weights delivers near-frontier coding scores. Guessing 744 billion made it look like a conventional large model competing on size. The real number makes it an efficiency story, and efficiency is what determines whether you can self-host the thing.
Inside Z.ai's GLM-5.3-Flash Architecture
The architecture is where this release earns its attention, and it is the part that the anonymous preview could not have communicated.
Sparse and linear attention, together
The model uses a hybrid attention design that combines linear attention for local dependencies with sparse attention for global context. Z.ai reports this cuts attention compute by 3.01 times and shrinks the key-value cache by 4.44 times relative to GLM-5.3. Those are the two numbers that decide whether a million-token context is a specification-sheet flourish or something you can actually serve.
Manifold-Constrained Hyper-Connections
The card also names a technique called Manifold-Constrained Hyper-Connections, presented as the mechanism that keeps training stable while scaling the sparse mixture. It is a research claim rather than a benchmark, and the honest position is that nobody outside Z.ai has yet reproduced it. It is now testable by anyone, which is the point of releasing weights.
IndexPool and the million-token window
The context window is 1,048,576 tokens with a maximum output of 131,072. A compression technique called IndexPool is credited with making that window affordable in memory terms. In practice the useful figure for most teams is the KV-cache reduction, because that is what sets how many concurrent long-context sessions a given GPU allocation will hold.
Reasoning is mandatory
One design decision has real operational consequences: thinking cannot be switched off. The API accepts effort levels of low, high or max, and rejects any attempt to disable reasoning entirely. That was, as it happens, the single most incriminating detail during the stealth period — and it means every request carries reasoning tokens whether the task needs them or not.
The specification at a glance
| Attribute | GLM-5.3-Flash | What it implies |
|---|---|---|
| Total parameters | 320 billion | Mid-weight by 2026 frontier standards |
| Active parameters | 18 billion per token | Low serving cost per request |
| Context window | 1,048,576 tokens | Whole-repository prompting is viable |
| Maximum output | 131,072 tokens | Long refactors in a single response |
| Training corpus | 30 trillion tokens, multimodal | Vision is native, not bolted on |
| Inputs accepted | Text, images, video, files | Screenshots and screen recordings work |
| Reasoning control | Low, high or max; cannot be disabled | Every call pays a reasoning premium |
| Licence | MIT | Commercial self-hosting permitted |
| Serving stacks | SGLang, vLLM, TokenSpeed, KTransformers | Standard open-source tooling |
The Benchmarks Z.ai Actually Published
With the weights came a scorecard, and it is a more modest document than the stealth-period hype implied.
Terminal Bench 2.1
On Terminal Bench 2.1, an agentic command-line benchmark, GLM-5.3-Flash scores 84.3. Claude Opus 4.8 sits at 85.0, Gemini 3.7 Flash at 85.8 and GPT-5.6 Terra at 87.4. So the model is fourth of four in that comparison — but the spread from top to bottom is 3.1 points, which is close enough that price becomes the deciding variable rather than capability.
DeepSWE and AutomationBench
On DeepSWE v1.1 the published figure is 63.4, against 46.2 for GLM-5.2. On AutomationBench it scores 48.8 against 26.2 for the same predecessor. Those generational jumps — 37 per cent and 86 per cent respectively — are the strongest numbers in the Z.ai release, and they are internal comparisons rather than competitive ones, which is the appropriate way to read them.
The house benchmark
Z.ai also publishes a proprietary measure, Code Bench v1.0, where GLM-5.3-Flash reaches 29.0 at maximum effort against 29.5 for Claude Opus 4.8. A vendor’s own benchmark showing a near-tie with the market leader is precisely the result you would expect a vendor’s own benchmark to show. Treat it as directional and run your own evaluation.
Vision and general reasoning
The model records 62.4 on OfficeQA Pro, a vision benchmark covering documents and dashboards, and 55.3 on Humanity’s Last Exam. Artificial Analysis places it at 57 on its Intelligence Index v4.1.1 at a cost of $0.045 per task on the discounted tier, which is where the efficiency argument stops being architectural and starts being financial.
Terminal Bench 2.1, four models compared
The 80 Per Cent Claim Z.ai Never Made
The most repeated number about this model during its anonymous week was wrong, and the correction travelled a fraction as far as the error.
Where the figure came from
Independent researcher Ben Davis ran DeepSWE against the stealth endpoint and reported 80 per cent Pass@1. Compared against roughly 65 per cent for Claude and roughly 52 per cent for the GPT-5.6 line, that looked like a Chinese open-weight model comprehensively beating the frontier. The screenshot went everywhere.
The detail in the footnote
It was a 10-task subset. Ten problems, on a benchmark that ships with 113. Eight successes out of ten produces exactly 80 per cent, and a single different outcome would have moved the figure by ten points in either direction. The sample was too small to distinguish a frontier model from a competent one.
The full run
Davis subsequently completed the whole 113-task benchmark and reported approximately 63 per cent Pass@1, describing the model as roughly on par with the mid tier of GPT-5.6 Sol. He flagged the full run as the representative result. When Z.ai published its own DeepSWE v1.1 figure of 63.4, it matched the corrected independent number almost exactly — which is a quiet vindication of both the researcher and Z.ai.
The claim that never got tested
A separate community report of 48 resolved tasks out of 50 on SWE-bench circulated in the same period and was never independently reproduced. Small test sets, unknown contamination, no methodology published. It should not have been repeated, and it was repeated widely.
What buyers should take from this
Model evaluation in public is now a race between a screenshot and a footnote, and the screenshot wins. If a benchmark number is going to influence a procurement decision, the sample size matters more than the score. This is the same discipline that any serious AI strategy applies to vendor claims, and it costs nothing to insist on.
The 80 per cent claim against the measured results
How the Z.ai Stealth Launch Actually Performed
Strip out the mystery and what remains is one of the largest unbranded product trials the industry has run.
Six days, 44 trillion tokens
The OpenCode dashboard on 26 August recorded 44 trillion tokens processed through the anonymous endpoint, across 13.12 million completed sessions and 503,000 unique users. That represented 10.6 per cent of the platform’s token traffic. Four days earlier the figure had been around 7.1 trillion tokens across roughly 134,000 users, so the great majority of that volume arrived in the final 48 hours as word spread.
The end of DeepSeek’s run
The model displaced DeepSeek at the top of the OpenCode leaderboard after a 56-day reign. Eight weeks is a long tenure in this market, and it ended without the challenger even having a name attached to it.
The serving telemetry
Reported performance during the preview was a median latency of around six seconds to first token, throughput in the region of 24 tokens per second, and effectively unbroken uptime. Throughput readings varied by platform and by reasoning effort, which is what you would expect from a model that cannot skip its thinking step.
The capacity claim
OpenCode advertised aggregate service capacity of 100 trillion tokens per day. The claim did not specify whether that counted input, output or cached reads, and it described headroom rather than delivered traffic. Against 44 trillion tokens across six days, the headroom was never seriously tested.
The preview scoreboard
| Metric | Figure at reveal | Reading |
|---|---|---|
| Tokens processed | 44 trillion in six days | Frontier-scale trial traffic |
| Completed sessions | 13.12 million | Roughly 26 per user |
| Unique users | 503,000 | Grew ~3.8x in the last two days |
| Share of platform tokens | 10.6 per cent | From zero, in under a week |
| Median latency | About 6 seconds | Slow, because reasoning is forced |
| Cost to the user | Zero for the preview week | The reason the volume existed |
How Developers Unmasked Z.ai Before the Announcement
The forensics deserve recording, because they worked, and because they are repeatable against the next anonymous endpoint.
The tokenizer fingerprint
Every model family tokenises text slightly differently. Testers fed identical strings — English, German, Chinese, source code, emoji, Korean, full-width Latin, Cyrillic — and compared token counts against known GLM behaviour. The counts matched with a fixed 75-token offset, consistent with a hidden system wrapper. No rival family came close to that correspondence.
Error code 1210
The decisive clue was an error message. The endpoint returned “[1210] This model always engages in thinking and cannot be disabled; please use low, high, or max” — a string matching Z.ai’s documented GLM-5.3 parameter contract exactly, right down to the numeric code. Labs share tokenizers by accident. They do not share error codes.
The reasoning-effort contract
The same low, high and max effort levels, with no option to disable reasoning, appeared in the public GLM-5.3 documentation. The API surface of an anonymous model is a fingerprint in itself, and this one had been published weeks earlier by the very lab the crowd was trying to identify.
The video-encoder tell
Video handling behaved identically to GLM-5V-Turbo, including how the endpoint rejected audio. Multimodal input was in fact the one capability that argued against a straight GLM-5.3 checkpoint, since the public GLM-5.3 announcement had not featured video — which is precisely why the community landed on a distinct Flash variant.
The evidence ledger
| Evidence | Strength before the reveal | Verdict after it |
|---|---|---|
| Tokenizer match with 75-token offset | Strong | Correct |
| Error code 1210 wording | Decisive | Correct |
| Low/high/max effort contract | Strong | Correct |
| Video encoder behaviour | Moderate | Correct |
| Nebius executive naming the model | Direct but retracted | Correct |
| Parameter estimate near 744 billion | Speculative | Wrong — it is 320 billion |
| 80% DeepSWE Pass@1 | Viral | Wrong — 63.4 on the full run |
Why Z.ai Ran the Whole Thing Anonymously
Anonymity is not modesty. It is a research method with a marketing budget attached.
Evaluation without a flag
An anonymous endpoint gets judged on output alone. No national-origin discount, no brand halo, no assumption that a Chinese checkpoint must be a distillation of somebody else’s work. For a lab whose principal commercial obstacle is scepticism about provenance, a week of blind testing is worth more than any launch blog.
Free load testing at frontier scale
Forty-four trillion tokens of real developer traffic, with real agentic loops and real long-context abuse, is a stress test no internal harness reproduces. The users paid nothing and supplied the most valuable input Z.ai could have bought: behaviour under load, plus a corpus of failure modes.
The marketing arithmetic
A one-week free window creates urgency by design — anyone who wants to test has to test now. The resulting speculation generated days of organic coverage across developer media, X threads and newsletters. Purchasing that attention through conventional channels would have cost a great deal more than a week of inference.
It is now a genre
This is not an isolated stunt. Through 2026 the pattern has repeated: Pony Alpha resolved to a GLM release in February, Hunter Alpha to Xiaomi’s MiMo line in March, Elephant Alpha to Ant Group’s Ling family in April, Owl Alpha to Meituan’s LongCat in late April. Ox Alpha is simply the largest and most successful instance so far.
The stealth-launch pattern in 2026
| Codename | Month | Lab it resolved to |
|---|---|---|
| Pony Alpha | February 2026 | Z.ai, then Zhipu / GLM-5 |
| Hunter Alpha | March 2026 | Xiaomi MiMo |
| Elephant Alpha | April 2026 | Ant Group Ling |
| Owl Alpha | Late April 2026 | Meituan LongCat |
| Ox Alpha | August 2026 | Z.ai / GLM-5.3-Flash |
What MIT-Licensed Weights From Z.ai Actually Change
An MIT licence on a near-frontier coding model is the part of this story with the longest half-life.
The licence in plain terms
MIT permits commercial use, modification, redistribution and private forks, with attribution and no warranty. There is no acceptable-use annexe, no monthly-active-user ceiling, no clause reserving the right to revoke. Compared with the bespoke community licences attached to many open-weight releases, it is unusually clean.
What self-hosting demands
Eighteen billion active parameters is the number that makes on-premises deployment plausible, but 320 billion total weights still have to be resident. SGLang, vLLM, TokenSpeed and KTransformers are all supported, which means the deployment path uses tooling most infrastructure teams already run rather than a proprietary runtime.
Trained and served on Chinese silicon
Z.ai reports deploying the model on domestic Chinese accelerators using a customised SGLang inference engine. Given the export controls in force, that is a strategic disclosure as much as a technical one: it asserts that a competitive model can be trained and served without US hardware.
The sovereignty argument
For organisations that cannot send source code to a third-party API — defence suppliers, regulated finance, anything under strict data residency — an MIT-licensed model you can run inside your own boundary is categorically different from a cheaper API. That is the case for evaluating it seriously, and it applies whether or not the benchmark scores impress you.
Where it fits alongside agents
The positioning statement mentions sustained agentic work, and that is the workload where a long context and a low active-parameter count compound. Teams already running AI agents against a metered frontier API will find the economics of a self-hosted alternative worth modelling, even if the accuracy trade is real.
Z.ai Pricing Against the Frontier Labs
The published rates are where the efficiency architecture converts into a commercial argument.
The headline numbers
The hosted API is priced at $0.15 per million input tokens and $0.50 per million output tokens, with cached input at $0.03 per million. Artificial Analysis measured an effective $0.045 per task on its Intelligence Index at a discounted tier. Against typical frontier pricing this is an order-of-magnitude difference, not a discount.
Why caching matters here
At $0.03 per million, cached input costs a fifth of fresh input. For agentic workloads that resend a large system prompt and repository context on every turn, cache hit rate becomes the dominant cost variable — more so than the headline rate. Any cost model built without it will overstate the bill substantially.
The reasoning premium
Set against that, thinking cannot be disabled, so every request generates reasoning tokens billed at the output rate. A short classification task that another model would answer in fifty tokens may cost several hundred here. The cheap rate and the forced reasoning partially cancel, and where they land depends entirely on your task mix.
Published price per million tokens
Who Z.ai Actually Is
The lab is better known by its old name, and its corporate history explains a good deal about how this launch was run.
A Tsinghua spinout
Z.ai, previously Zhipu AI, was founded in 2019 out of Tsinghua University’s Knowledge Engineering Group by professors Tang Jie and Li Juanzi. Zhang Peng is chief executive. The GLM series has been its flagship line throughout, and Z.ai is one of the small group of Chinese labs shipping genuinely open weights rather than API access alone.
The Hong Kong listing
The company listed on the Hong Kong Stock Exchange on 8 January 2026 under the code 2513.HK, at an offer price of HK$116.20 per share, raising roughly HK$4.17 billion at a valuation near $7 billion. The retail tranche was oversubscribed more than a thousand times. Around 70 per cent of proceeds were earmarked for model research through 2028.
The financial picture
Revenue for 2025 was RMB 724 million, about $105 million, up roughly 132 per cent year on year. The net loss was RMB 4.7 billion. Pre-IPO backing of about $1.5 billion came from Alibaba, Tencent, Ant Group, Meituan, Xiaomi and Saudi Aramco’s Prosperity7 Ventures. Those are the numbers of a company buying market share, and free preview weeks are how you buy it.
The entity list
Z.ai was added to the US entity list in January 2025 on military-modernisation grounds, restricting its access to American technology without a licence. That constraint is the backdrop to the domestic-silicon disclosure, and it is a live consideration for any Western buyer weighing the hosted API against self-hosted weights.
Company facts
| Item | Detail |
|---|---|
| Founded | 2019, Beijing, from Tsinghua University’s Knowledge Engineering Group |
| Founders and chief executive | Tang Jie and Li Juanzi; Zhang Peng leads Z.ai |
| Listing | Hong Kong, 8 January 2026, ticker 2513.HK |
| Offer price and raise | HK$116.20 per share, about HK$4.17 billion |
| 2025 revenue and loss | RMB 724 million revenue, RMB 4.7 billion net loss |
| Pre-IPO investors | Alibaba, Tencent, Ant Group, Meituan, Xiaomi, Prosperity7 |
| US entity list | Added January 2025 |
Risks and Open Questions About Z.ai and This Release
Four things about this release remain genuinely unresolved, and none of them are settled by the benchmark table.
The retention contradiction
During the preview the two routes had different terms. OpenCode Zen promised zero-day retention with no training use. The anonymous provider behind the OpenRouter listing retained prompts and completions, stating only that they would not be used for training. That is not the same thing as zero retention, and the marketing language did not distinguish them. Anyone who pasted proprietary code into the free endpoint should assume it persisted somewhere.
Provenance of the training data
A 30-trillion-token multimodal corpus is described but not itemised. No dataset disclosure accompanied the weights, which is standard practice across the industry and no less unsatisfactory for being standard. Post-training methods, including any use of reinforcement learning on coding trajectories, are likewise unspecified.
Supply and continuity
A lab on the entity list, serving on domestic accelerators, carries a different continuity profile from a US hyperscaler. The MIT licence mitigates this considerably — weights already downloaded cannot be withdrawn — but the hosted API and any future checkpoint sit on the other side of an export-control regime that has tightened repeatedly.
Jurisdiction and governance
For a UK or EU organisation, using the hosted API means data crossing into Chinese jurisdiction. Self-hosting the MIT weights sidesteps that entirely, which is the single strongest argument for treating this as a weights release rather than an API launch. The governance question is where the inference runs, not who trained the model.
What Z.ai Means for Technology Buyers
The practical response is neither to adopt on the strength of a leaderboard nor to dismiss on the strength of a flag.
Evaluate on your own tasks
Terminal Bench and DeepSWE measure generic software-engineering ability. They do not measure your codebase, your languages, your review standards. A fifty-task internal evaluation on real tickets will tell you more in a day than three months of leaderboard-watching.
Model the cost properly
Build the cost model around cache hit rate and reasoning tokens, not the headline per-million rate. Run the same workload at low and at max effort and compare quality against spend; the gap between the two is often smaller than the price difference. Serving economics is also where the hardware question resurfaces, as we covered in our piece on inference chip benchmarks.
Decide the hosting question first
Hosted API and self-hosted weights are different products with different risk profiles. Decide which one you are actually assessing before you start, because the answer determines whether jurisdiction, entity-list exposure and data residency are live concerns or irrelevant ones.
Keep the abstraction layer
Whatever you choose, route through a gateway rather than hard-coding one provider. The lesson of a stealth model reaching 10.6 per cent of a platform’s traffic in six days is that the leaderboard changes faster than a procurement cycle completes.
A buyer’s checklist
| Question | Why it decides the outcome |
|---|---|
| Hosted API or self-hosted weights? | Sets every jurisdiction and residency answer that follows |
| What is your cache hit rate? | Cached input is one fifth the price of fresh input |
| Can you tolerate forced reasoning? | Short tasks pay a premium that cannot be switched off |
| Do you have GPU capacity for 320B weights? | Active parameters are low; resident weights are not |
| Did anyone use the free preview with real code? | Retention terms differed by route during that window |
| Is your integration provider-agnostic? | Six days was enough to reorder a leaderboard |
Frequently Asked Questions About Z.ai and Ox Alpha
Which lab was behind Ox Alpha?
Z.ai, the Beijing lab formerly known as Zhipu AI and the maker of the GLM series. It confirmed authorship on 26 August 2026, first through Bloomberg and then in wider coverage, after the model had spent six days on OpenRouter and OpenCode without an owner.
What is Ox Alpha’s real name?
GLM-5.3-Flash. The weights were published on Hugging Face as zai-org/GLM-5.3-Flash on the evening of the confirmation, under an MIT licence.
How big is the model?
320 billion total parameters with 18 billion active per token, in a mixture-of-experts configuration trained on a 30-trillion-token multimodal corpus. Community estimates during the anonymous week had guessed roughly 744 billion total, more than double the real figure.
Did it really beat Claude and GPT on benchmarks?
Not on the published numbers. It scores 84.3 on Terminal Bench 2.1 against 85.0 for Claude Opus 4.8 and 87.4 for GPT-5.6 Terra, and 29.0 against 29.5 for Opus 4.8 on Z.ai’s own Code Bench. The viral 80 per cent DeepSWE result came from a 10-task subset; the full 113-task run returned approximately 63 per cent, matching the published 63.4.
Is it free?
It was free during the preview week that ran from roughly 20 to 27 August 2026. The hosted API is now priced at $0.15 per million input tokens, $0.50 per million output tokens and $0.03 per million cached input tokens. The weights themselves are free to download and run under MIT.
Can I run it on my own hardware?
Yes. The MIT licence permits commercial self-hosting, and the model is supported by SGLang, vLLM, TokenSpeed and KTransformers. You need enough memory for 320 billion resident weights, though only 18 billion are active per token.
Why did Z.ai keep it anonymous?
Blind evaluation, free load testing at scale and cheap attention. An unbranded endpoint is judged on output rather than origin, and 44 trillion tokens of real developer traffic is a stress test no internal harness can reproduce.
Was my code retained if I used the free preview?
Possibly. OpenCode Zen advertised zero-day retention, but the anonymous provider behind the OpenRouter listing retained prompts and completions while stating they were not used for training. Treat anything sent through that route as having persisted.
How did developers work out who built it?
Tokenizer probes matching GLM behaviour with a fixed 75-token offset, an error string carrying code 1210 and Z.ai’s exact reasoning-effort wording, video handling identical to GLM-5V-Turbo, and finally a Nebius executive naming the model in a post that was quickly taken down.
References and Further Reading
Surprise: Z.ai is the AI lab behind the mysterious Ox Alpha model
GLM-5.3-Flash model card on Hugging Face
GLM-5.3-Flash official documentation
Z.ai launches GLM-5.3-Flash under MIT license
Mystery Ox Alpha model revealed to be from Chinese lab Z.AI
Z.AI reveals Ox Alpha is GLM-5.3 Flash: the benchmark table
Ox Alpha confirmed as Z.AI’s new GLM — what we know
Ox Alpha unmasked: GLM-5.3-Flash clocks 44T tokens
Is Ox Alpha GLM-5.3 Flash? Evidence and gaps
Stealth frontier-level model Ox Alpha appears on OpenRouter and OpenCode
Ox Alpha: a mysterious new AI model aims to win developers over
Zhipu AI (Z.ai) in 2026: models, IPO, funding and strategy
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.