Union Alpha on OpenRouter: Tests and Reviews

TNG's probe tests found timeouts varying by domain, widely fluctuating token counts, inconsistent knowledge cutoffs and alignment, and a 50x latency spread, leading it to conclude Union Alpha is a multi-backend auto-router.

A roundup of the anonymous model's selling points: agentic coding focus, image support, 262K context, no training on user data, and one free week on OpenCode and OpenRouter.

A user shares a screenshot saying Union Alpha's output style matches GLM, suspecting the anonymous model is actually Zhipu's.

mehulmpt says Union Alpha is unlikely to be a GLM model, since Ox Alpha already taught GLM how popular such releases can be, and it looks more like another lab learning its inference-scale lesson.

The author cites zero data retention, a 262K context window, EU/Spain ties, release timing and Mistral's recent funding as reasons Union Alpha may be Mistral.

After hours of analysis, the author says Union Alpha's text tokenizer was modified and its vision tokenizer matches GLM-5.3 Flash, concluding it is GLM-5.5.

A user reports Union Alpha's free week brought frequent outages, 22 tok/s and cut-off responses, even after the provider said it had added capacity.

A video puts OpenRouter stealth model Union Alpha against a locally run DeepSeek V4.1 Flash on the same real tasks, and the results differ notably.

Comparing Union Alpha with other stealth models like 0x alpha and omen alpha, the post notes it is the only one with an X account, was extremely slow at first before the team fixed it, is free for only a week, and may not be a GLM or Kimi.

A user who retested Union Alpha says its 3D and SVG output is far worse than the Flash models, and speculates it could be a Mistral model with 256K context if it comes from Spain.

OpenCode says several improvements were made for Union Alpha and invites developers who hit errors or odd behavior to try it again in the harness.

A user says Union Alpha built a 3D demo in 10 minutes and shared a screen recording, noting that with more time it might reach Astra/Opus-level graphics but they didn't want to wait.