Ox Alpha Offers a Million Tokens, but Not a Name

Ox Alpha arrived with a concrete promise: a 1,048,576-token context window and free access for one week. But the model appeared on OpenRouter under the provider name Stealth, with no developer claiming ownership and no disclosed architecture, formal benchmarks or post-preview price.

Inputs and workflow controls

A context window is the amount of material a model can process at once. Ox Alpha’s listed limit is 1,048,576 tokens, a scale the source compares with an entire codebase or several full-length novels in a single prompt. That capacity is relevant to work that needs a model to hold extensive source material without splitting it into smaller requests.

The listing also says Ox Alpha accepts text, image and video inputs. It supports tool and function calling, which lets a model invoke defined software actions, alongside structured JSON output for machine-readable responses. Those interfaces make the preview more than a long-context chat demonstration, even though they reveal little about the underlying system.

The identity case is suggestive, not settled

Community researchers have tied Ox Alpha to Z.ai’s GLM family using tokenizer fingerprinting and API response patterns. A tokenizer is the system that breaks text into tokens, so its behavior can offer a recognizable technical signature. The association is not a confirmation: no ownership claim or formal identification had been reported.

There is another possibility in the speculation. Xiaomi’s MiMo team has been floated as a secondary theory because it has used similar stealth-release tactics. The account also describes Ox Alpha as the fifth anonymous OpenRouter model linked to Chinese labs, and says earlier drops including GLM-5 and MiMo-V2-Pro were later claimed after their preview periods ended.

A window size is not a performance verdict

The missing disclosures set the limit on what the launch establishes. A large window describes how much input the model can take in, not how reliably it reasons over that input or how it compares on specific tasks. Architecture details and formal benchmarks had not been disclosed, and neither had the price once the free period ends.

The provider did state that prompts submitted during the preview would not be used for training. That is a meaningful operating commitment for anyone evaluating the model with real material, but it sits alongside the larger unanswered question of who is operating the service. For now, Ox Alpha is testable software with a substantial advertised capacity, not a fully characterized model release.