Ox Alpha
A new “stealth” LLM called Ox Alpha, available for free via OpenRouter, is drawing attention because its creator is undisclosed, yet prompts and completions are retained by the provider for analytics while allegedly not being used for training. Users speculate it is a Chinese-origin GLM 5.x variant or a mix of models, probing it with politically sensitive queries like Tiananmen Square, Taiwan, and Tibet to infer censorship patterns and compare its guardrails to those of US labs. While many find it surprisingly capable—especially for creative work and at high speed—others voice concern over opaque provenance, privacy risks, licensing implications, and the broader trend toward less transparency in frontier AI models.
Model identity & behavior
- Many think Ox Alpha is a Chinese model, likely GLM 5.x (possibly 5.3, maybe a vision-capable or “Air” variant); others suggest it might be DeepSeek or a Xiaomi model.
- Some claim the behavior and chain-of-thought style strongly resemble GLM, and that prompt/response patterns match GLM 5.3 outputs.
- A few suggest it might be a model router using multiple backends, based on inconsistent answers across users.
Guardrails, censorship & politics
- Users probe the model with “China 1989 / Tiananmen Square,” Tibet, and Taiwan questions as a canary for Chinese censorship.
- Reports conflict: some say it refuses Tiananmen content but gives offensive cyber help; others say it freely describes Tiananmen and criticizes the CCP.
- There’s debate over whether censorship is at the model-training level or at the serving endpoint, and whether future systems will bake censorship into weights.
- Several compare Chinese political redlines (e.g., Tiananmen, Tibet) with Western labs’ refusals on drugs/malware/biology, arguing both are topic-specific safety limits.
Privacy, data use & “stealth” concerns
- Ox Alpha is a “stealth model”: OpenRouter knows the provider but does not reveal it; prompts/completions are retained but (claimed) not used for training.
- Some are deeply skeptical; they doubt the “no training” promise is verifiable and suspect legal hair-splitting (e.g., using data for RLHF/analytics).
- Others argue this is standard in the industry: if you care about secrecy, don’t send anything sensitive to “mystery endpoints.”
- There’s broader concern that all major labs train on whatever they can, despite official policies, driven by data scarcity and competitive pressure.
- Some prefer Chinese open-weight models because they can self-host or choose trusted hosting with strict retention policies.
Capabilities, use cases & UX
- Several find Ox Alpha very strong on creative/“soft” tasks, sometimes beating well-regarded models, though weaker in visual reasoning and frontend/CSS work.
- Speed impressions are mixed: some call it “suspiciously fast,” others say GLM 5.3 is usually slower than top US models in real tasks.
- Knowledge cutoff is estimated around mid‑2025.
- Free access and low rate limits are seen as a way for labs to smoke-test models, gather usage analytics, test infra/capacity, and gain market foothold without PR risk.