Ox Alpha finally dropped the mask: after six days of guessing, it was Zhipu's GLM-5.3-Flash

Conclusion

The mystery model that the internet spent six days guessing has finally dropped the mask.

On August 26, Zhipu's Z.ai confirmed that the model it tested anonymously on OpenCode and OpenRouter as ox-alpha was GLM-5.3-Flash. Its model card, parameter details and weights are now on Hugging Face under the MIT License.

This is more than a funny story about internet detectives being right. During the anonymous run, Ox Alpha topped OpenRouter's popularity ranking. Bloomberg reported that its usage at one point exceeded DeepSeek by more than two times. After the reveal, three other numbers made the story more consequential: 320B total parameters, 18B active parameters and a 63.4 DeepSWE result.

But "approaching Opus 4.8" is Z.ai's product positioning, not proof that the model has beaten Claude across every Agent workload. The durable news is that a Chinese open-weight model removed its brand and persuaded developers around the world to vote with real usage before they knew who made it.

Six anonymous days reversed the normal launch sequence

When Ox Alpha appeared around August 20, it had no company logo, parameter table or developer identity. It offered a free endpoint, a million-token-class context window and an invitation to throw real repositories at it.

The normal launch sequence begins with vendor benchmarks, moves through media coverage and ends with developers trying the product. Z.ai reversed it: hide the name, provide free access, let people hammer it through OpenCode and OpenRouter, then reveal the identity and weights after usage and speculation had accumulated.

That is especially effective for a Chinese model. If the Zhipu name had appeared on day one, some overseas users might never have clicked. Removing the brand forced them to decide first whether the model could do the work.

Developers really did run a DNA test

Before the official reveal, community black-box work had already narrowed the model to the GLM-5 family. The strongest evidence did not come from asking the model who it was. Researchers counted how its tokenizer split Chinese, Japanese, emoji, code and unusual whitespace.

Across 44 discriminating strings in a public evidence repository, Ox Alpha matched the GLM-5-generation tokenizer 44 out of 44 times. GLM-4.x missed two emoji cases. API errors, reasoning controls and context limits pointed in the same direction.

That fingerprint could identify the GLM-5 generation, not an exact checkpoint, because GLM-5, 5.1 and 5.2 share important tokenizer behavior. The official confirmation supplied the last missing step: Ox Alpha did not merely resemble GLM. It was the pre-release anonymous test of GLM-5.3-Flash.

Three hundred twenty billion parameters, but only 18B wake up at once

GLM-5.3-Flash is a mixture-of-experts model with 320B total parameters and 18B activated for each token. That does not mean an 18B model is pretending to be 320B. It means the system has a large pool of experts and routes each token through a small subset.

Z.ai says it trained a new base model rather than simply trimming GLM-5.3. This is the first natively multimodal model in the GLM-5 series. It combines sparse and linear attention, uses mHC and was pretrained on a 30-trillion-token multimodal corpus.

That explains the Flash positioning. The target is not only the highest score, but the amount of compute required for each Agent task.

A 63.4 is strong, but a vendor table is not "Opus destroyed"

Z.ai reports 63.4 on DeepSWE v1.1, up from 46.2 for GLM-5.2. It reports 48.8 on AutomationBench versus 26.2, and the model card lists 84.3 on Terminal Bench 2.1.

The configurations are documented, which is better than a promotional ranking with no method, but these remain vendor-published results. DeepSWE used mini-swe-agent, a six-hour timeout and 400K context. Terminal Bench used a Claude Code harness. A different harness, concurrency policy, time budget or cost ceiling can change the result materially.

Z.ai's careful wording is that GLM-5.3-Flash approaches Claude Opus 4.8 on coding and agentic benchmarks. We do not have enough evidence to upgrade that to a claim of comprehensive superiority. The more useful question is whether it can be good enough in a developer's repository, tools and time budget at a much lower cost.

Price is the real weapon in the Flash label

The official model card positions GLM-5.3-Flash as outperforming GLM-5.2 across several workloads at roughly one-tenth the price while approaching the Opus 4.8 range in coding and agents. Z.ai also says the model reached 57 on Artificial Analysis Intelligence Index v4.1.1 at a discounted $0.045 per task, an intelligence range that previously cost about ten times more per task.

"One-fortieth the price of Opus" is an excellent headline number, but it depends on launch discounts, task length and billing assumptions. It should not be treated as a permanent ratio for every workload. The safer conclusion is that GLM-5.3-Flash is trying to become the default by pairing near-high-end Agent capability with a sharply lower task cost.

The anonymous traffic ran on domestic chips, according to Z.ai

Z.ai states in its launch blog that Ox Alpha became the most popular model of the week on OpenCode and OpenRouter and that "all of this traffic" was served on Chinese AI chips. It says a large-scale domestic-chip cluster, high-bandwidth interconnect and hardware-optimized serving stack handled the past week.

That matters because this was not a small benchmark run. It was a free, high-traffic test with developers around the world. The evidence is still Z.ai's own statement. Public material does not name the chip model, cluster size, per-node efficiency, power or failure rate. We can report that Z.ai says domestic chips carried the traffic; we cannot infer a supplier or compare efficiency with Nvidia.

This was not only an identity claim; the weights really arrived

When Bloomberg revealed the identity, Z.ai was still promising to open the weights that evening. The Hugging Face repository zai-org/GLM-5.3-Flash is now live with an MIT license and deployment paths for vLLM, SGLang, TokenSpeed and KTransformers.

Open weights move the story into its next phase. The community can inspect the architecture, quantize it and rerun tasks across different hardware. But 320B total parameters still means this is not an 18B model that an ordinary gaming PC can load effortlessly. Active parameters shape much of the compute per token; total weights still determine storage and memory pressure.

Our judgment

"Ox Alpha unmasked" is more meaningful than "Zhipu released another model" because it adds evidence that a vendor cannot manufacture alone. Developers used the system heavily before they knew the brand.

That does not make it the world's strongest model. It does support a more practical product thesis: many everyday Agent jobs do not need the most expensive model on every call. If independent testing keeps GLM-5.3-Flash around the 60%+ DeepSWE tier, it may compete less for the strongest-model crown than for the much larger pool of default daily work.

The anonymous-release tactic will probably be copied. Let real usage and community opinion form before revealing the brand, especially when geography creates bias. The cost is a new debate over free-period data policy, reliability and marketing transparency.

What still needs watching

The most useful evidence over the next 24 to 72 hours will be independent reruns after the weights became available. On the same DeepSWE version, harness, cost ceiling and time limit, how large is the gap to Claude, GPT, DeepSeek, Kimi and Qwen?

The second question is production API behavior: throughput under load, time to first token, cache-hit rates, errors and throttling. Users reported congestion and long stalls during the free anonymous period. Average benchmark scores cannot hide those experiences.

The third question is hardware. Which domestic chips, cluster scale and power budget carried the anonymous traffic, and can the serving stack sustain commercial traffic at the announced economics?

Sources and evidence boundary

• Primary model card and weights: Z.ai's Hugging Face repository supports 320B/18B, the MIT license, native multimodality, architecture and deployment paths;
• Primary launch material: Z.ai's blog supports the Ox Alpha identity, vendor benchmark results, task-cost claim and domestic-chip serving statement;
• Reputable media: Bloomberg supports Z.ai's confirmation, the OpenRouter lead and usage exceeding DeepSeek by more than two times;
• Reproducible community forensics: a public GitHub repository supports the 44/44 tokenizer fingerprint and the boundary between generation attribution and exact-checkpoint attribution;
• The 63.4, 48.8, 84.3 and $0.045-per-task figures are vendor-published results, not independent AI Radar reruns.