Tests link Ox Alpha to Zhipu GLM API routes and error codes

• A malformed Ox Alpha request exposed an internal com.wd.paas.api.domain.v4.chat.ChatCompletionRequest class and a route that chetaslua's trace tied to Zhipu's serving stack.
• Matching 1214 validation errors across Ox Alpha and three Z.ai GLM endpoints, alongside a divergent same-weights DeepInfra control, are documented in chetaslua's control test.
• The preview arrived as a free 1M-context multimodal model in OpenRouter's launch post, while its public model page still describes an anonymous third-party operator.
• A Roboflow object-detection comparison found Ox Alpha at 4.8% mAP@50 versus Gemini 3.7 Flash at 99.9% on the same banana-tree image, according to skalskip92's test.

An OpenRouter model page says the unnamed provider, rather than OpenRouter, operates Ox Alpha and retains prompts and completions without using them for training. Z.ai's Chat Completion reference publishes the same /paas/v4/chat/completions path that surfaced in the trace. The useful clue came from fault handling, not chatbot vibes: according to chetaslua's gateway notes, OpenRouter sanitized upstream errors while OpenCode let malformed fields through.

A string supplied as top_p to OpenCode's x-preview-f-free route reportedly reached an upstream Java service and returned the internal com.wd.paas.api.domain.v4.chat.ChatCompletionRequest class name. The path in the trace was /api/paas/v4/chat/completions.

Z.ai's documentation specifies POST /paas/v4/chat/completions for its chat-completion API, including text and vision request schemas. That endpoint match pairs a publicly documented contract with the internal namespace in the malformed-request trace.

Chetaslua then published a curl reproduction for the malformed top_p request and said the behavior depended on an OpenCode pass-through bug that could be patched.

According to chetaslua's gateway notes, OpenRouter rewrote provider metadata and replaced errors with ERROR, while OpenCode forwarded malformed payloads. A role: "wizard" probe produced code 1214, and an out-of-range temperature produced 1210 plus Chinese-language validation text.

The key control was host-specific. Ox Alpha and OpenRouter instances of z-ai/glm-5.3, glm-5.2, and glm-5v-turbo all returned {"code":"1214","message":"Incorrect role information"} for the invalid role, while glm-5.2 on DeepInfra returned a Pydantic error, as chetaslua's control test records.

A separate AiBattle_ side-by-side screenshot showed the same top_p validation wording from the Zhipu international endpoint and Ox Alpha via OpenCode. Together, those tests attribute the shared behavior to the serving layer rather than treating a tokenizer match as the entire case.

In a chetaslua follow-up, 30 tokenizer probes across 14 writing systems, emoji, code, and SQL matched GLM-5.3. A later chetaslua post described Ox Alpha's vocabulary as GLM with two Chinese web-crawl tokens removed and no additions, calling it a GLM-5-generation post-training variant.

The public GLM-5.3 documentation describes a text-only model with a 1M-token context and 128K maximum output. OpenRouter's announcement describes Ox Alpha as taking text, image, and video. The public materials leave the exact checkpoint, vision stack, and eventual release name unpublished.

In skalskip92's Roboflow comparison, Ox Alpha detected 22 of 43 expected banana trees and scored 4.8% mAP@50. Gemini 3.7 Flash detected 44 against the same 43 ground-truth objects and scored 99.9%.

The wider skalskip92 thread also covered visually similar objects, dense scenes, hand-drawn diagrams, and bounding-box placement. It is a capability comparison on a particular evaluation suite, separate from the API-route attribution, but it sharply distinguishes the two models' observed vision profiles.

OpenCode's preview announcement advertised Ox Alpha with “Zero Data Retention.” The OpenRouter model page says prompts and completions are retained by the provider, while excluding them from training use.

The two public descriptions use different data-policy language: one names zero retention, while the other explicitly describes provider retention and a no-training condition.

kevinkern reported that Ox Alpha sometimes returned an empty response with zero output tokens and an unknown finish reason, which OpenCode treated as completed while Pi retried. An OpenCode issue describes the same silent-stop behavior for an empty finish: unknown response, observed there with a different gateway model.

OpenCode later said version 1.18.21 would automatically retry unknown stops. That implementation detail became part of Ox Alpha's hands-on record because gateway completion semantics could end an agent run before the model produced an answer.