On March 11, two AI models appeared on OpenRouter without announcement, attribution, or explanation. One week later, they’ve processed over 160 billion tokens, sparked a detective hunt across the AI community, and remain completely anonymous.
Meet Hunter Alpha and Healer Alpha, the stealth releases that may represent China’s next frontier models testing in the wild.
The Specs That Sparked a Mystery
Hunter Alpha claims to be a trillion-parameter model with a one-million token context window. It handles text and image input, outputs text only, and is marketed for “agentic use” with strengths in “long-horizon planning, complex reasoning, and sustained multi-step task execution.”
Healer Alpha is billed as an “omni-modal” model accepting text, image, audio, and video inputs. Its context window is smaller at 262K tokens, but it brings multimodal perception that Hunter lacks.
Both models are completely free. Zero cost per million tokens for input or output. The catch: “All prompts and completions for this model are logged by the provider and may be used to improve the model.”
That’s not charity. That’s crowdsourced testing at scale.
The DeepSeek Theory
The circumstantial evidence pointing to DeepSeek is substantial.
When Reuters tested Hunter Alpha, the chatbot described itself as “a Chinese AI model primarily trained in Chinese” with training data extending to May 2025. That cutoff matches DeepSeek’s own chatbot exactly.
The specifications align with long-rumored DeepSeek V4 capabilities: trillion-parameter scale, million-token context, multimodal support. After five missed release windows for V4 (we covered the delays last week), stealth testing through a third-party platform would explain both the silence and continued development.
According to Chinese tech outlet Whale Lab, DeepSeek V4’s official launch is now expected in April. Hunter Alpha could be generating the real-world usage data needed for final tuning.
The Zhipu Theory
But there’s a pattern here that points elsewhere.
The same anonymous OpenRouter provider account previously released “Pony Alpha,” which later turned out to be GLM-5 from Zhipu AI. In February, Zhipu confirmed Pony Alpha belonged to its GLM-5 system.
If the same playbook is repeating, Hunter Alpha could be Zhipu’s unreleased GLM-6 text model, and Healer Alpha could be GLM-5V or a new multimodal variant. The parameter jump to trillion scale would represent Zhipu’s attempt to match DeepSeek’s announced capabilities.
Community analysis remains split on architectural signatures. Some developers point to reasoning patterns consistent with DeepSeek’s R1 methodology. Others identify structural differences suggesting a separate origin.
How They Actually Perform
Early benchmarks are mixed. Ethan Mollick’s tests on March 12 showed Hunter Alpha performing “only okay” on reasoning challenges like the Lem Test and Sparks TiKZ unicorn challenge.
Testers note stronger censorship than previous DeepSeek models and weaker math performance. For a model positioned as a frontier release, these gaps are notable.
But the agentic use case may be the point. Hunter Alpha is explicitly designed for frameworks like OpenClaw, the open-source agent that recently triggered a security crisis with over 21,000 exposed instances. Long-horizon planning and instruction-following matter more than benchmark scores for sustained autonomous operation.
The Stealth Testing Strategy
Anonymous releases on third-party platforms solve multiple problems for Chinese AI labs.
Scale without infrastructure. OpenRouter handles distribution, API management, and billing. The model creator gets worldwide exposure without building global infrastructure.
Real-world feedback before commitment. Production usage generates data that internal testing can’t replicate. Logging all prompts and completions provides exactly the edge cases and failure modes needed for tuning.
Deniability. If the model underperforms or generates problematic outputs, there’s no brand damage. The release never officially happened.
Regulatory distance. Models distributed through US-based platforms face different scrutiny than direct exports from Chinese companies.
This isn’t new. It’s just usually done quietly. The Hunter/Healer releases are notable for their scale and ambition, not their strategy.
What This Means
Whether these models come from DeepSeek, Zhipu, or somewhere else entirely, the implications are the same.
Chinese AI labs are now confident enough to deploy trillion-parameter models for public testing. The infrastructure exists to do this anonymously through international platforms. And the community is eager enough to process 160 billion tokens in a week on models with unknown provenance.
For the local AI community, this is a preview. When these models eventually get official releases and open weights, they’ll likely be available for self-hosting. Million-token context on consumer hardware could become reality in months.
For privacy, the implications are more complex. Free models that log all prompts attract users who might otherwise choose paid services with better data policies. The tradeoff (frontier AI capability in exchange for your queries becoming training data) is explicit but easy to overlook.
What We’re Watching
Both DeepSeek and Zhipu have ignored requests for comment. OpenRouter hasn’t disclosed the provider’s identity. The models continue running, logging tokens, and refining themselves on user data.
April is the rumored window for DeepSeek V4’s official launch. If Hunter Alpha vanishes before then and V4 shows similar characteristics, the connection will be confirmed retroactively. Until then, the AI community has its most intriguing mystery of 2026.
Try them if you want frontier-level capability for free. Just remember: if the product is free and logs everything, you’re the product.