Error 1210 Just Exposed Zhipu's Secret AI

OpenCode posted a teaser on Wednesday. An anonymous multimodal model called Ox Alpha would go live next week, with 1M context length and support for images and video. The comments below the post split into two camps. One camp dug through Xiaomi's MiMo V2.5 Pro technical report, spotted the same bullet points—1M context and visual capabilities—and concluded that Ox Alpha was a pre-release test version of MiMo V3. The other camp noticed that Ox Alpha's response structure on OpenRouter looked unusually similar to Zhipu's GLM-5.3, leaning toward Zhipu.

Neither camp offered evidence. That's understandable: public information about an anonymous model is scarce, just an API endpoint and a brief call document. But the API itself can serve as evidence. Someone asked for Ox Alpha's API address in an OpenCode developer group, got it, and ran a comparison.

The experiment: prepare 25 prompts covering writing, mathematical reasoning, code generation, and general knowledge. For each, send requests to both Ox Alpha and GLM-5.3's public APIs and record the raw token counts. The result: a fixed difference. Ox Alpha consumed exactly 75 more tokens than GLM-5.3 every single time.

At the tokenizer level, a fixed offset points to the same underlying vocabulary with a constant-length prefix placeholder added. That's the same "root" with a different "affix," not a coincidental approximation between two independently trained models. If Ox Alpha were based on MiMo, Qwen, or Llama, its tokenization would produce fluctuating differences against GLM-5.3, not a fixed one. A fixed offset is the fingerprint of the same tokenizer.

Next, someone compared the API parameters. Both models force-enable reasoning mode. Both offer only three thinking-strength options: low, high, max. Both default to max. The defaults for temperature, top_p, and repeat_penalty were checked item by item and were identical down to the second decimal. If one default differed by 0.02, you could chalk it up to version tuning. But identical down to the last digit points more likely to the same backend config served through different front-end gateways.

The final data point came from an erroneous operation. Someone passed a malformed system prompt parameter to Ox Alpha. The API returned error code 1210 with a description. He screenshotted it and posted it in the developer group. A developer who had used Zhipu's open platform recognized the number: Zhipu's API documentation has a dedicated page explaining the trigger conditions and resolution for 1210. No other major platform assigns that same number to that type of error.

A model's inference interface, parameter defaults, and error-handling middleware all point to the same company. This goes far beyond coincidence.

Where does this model's multimodal capability come from? It accepts images and video, yet the public impression of Zhipu still centers on the pure-text and coding abilities of the GLM-4 and GLM-5 series. In reality, Zhipu has released two to three vision models per year since 2025:

• Mid‑2025: GLM‑4.1V‑Thinking, the first to output chain‑of‑thought on vision tasks.

• Late 2025: GLM‑4.5V strengthened GUI recognition; GLM‑4.6V improved temporal understanding of video clips.

• April 2026: GLM‑5V‑Turbo, which plugged the vision encoder into a coding workflow so the model could modify code by looking at screenshots.

Ox Alpha isn't Zhipu suddenly pivoting to multimodal. It merges GLM‑5.3's long‑text backbone with GLM‑5V's vision trunk.

One question remains: why did Zhipu release it anonymously?

Last year, DeepSeek‑V4‑Pro‑0813's launch announcement had already gone up on the official website, but third‑party benchmark results didn't match what Zhipu's internal Harness had produced. The announcement was taken down, and the follow‑up was low‑key. The impact wasn't about the technical issue itself; it was about the "posted then pulled" move, which erodes trust in a way that's harder to manage than a low score.

Zhipu ran a similar pilot in March. An anonymous model called Pony Alpha appeared on OpenRouter and was confirmed two weeks later as a test version of GLM‑5. By the time the community noticed, the model had already completed its final real‑user stress test, and Zhipu promptly released the official GLM‑5.

Ox Alpha follows the same playbook: no brand label, no product page, just an endpoint, letting real users hit it with production‑grade prompts. The issues uncovered in such tests are often invisible to internal evaluation. For example, a malformed JSON that parses incorrectly in a long context, or a video frame‑sampling step that overflows at a certain resolution. These fixes don't require a major version bump, but if they surface under a branded label, the explanation cost far outweighs the fix itself.

As for MiMo V3, there's one unexplained clue. Someone ran a small test between Ox Alpha and MiMo V2.5 Pro and found that both could recognize a specific token that GLM‑5.3 cannot handle. This suggests that MiMo and Ox Alpha may have training‑data overlap, share some portion of weights in an intermediate representation layer, or both data‑cleaning pipelines ingested the same external source. What can be confirmed right now is only this: Ox Alpha's error code points to Zhipu, and that piece of evidence outweighs all the other clues.

Next week, Ox Alpha's interface will be fully open. At that point, you can run a set of adversarial tests designed for Zhipu models, for instance, use a prompt format that GLM‑5.3 falls for and see if Ox Alpha reacts the same way. That result will be closer to the truth than any of today's speculation. One developer in the group said: right now I'm more curious about what the number 1210 actually corresponds to inside Zhipu. Maybe it's a gateway error‑table sequence number. Or maybe it's someone's birthday.