Hunter Alpha on OpenRouter - Is This DeepSeek V4?
A 1-trillion-parameter model called Hunter Alpha appeared anonymously on OpenRouter on March 11. Developers say it's DeepSeek V4 in disguise. The signals are strong but the precedent cuts both ways.
6 min read
"I am a Chinese AI model primarily trained in Chinese," Hunter Alpha said when a developer asked who built it. It wouldn't say more.
A 1-trillion-parameter AI model appeared on OpenRouter on March 11, fully free, with no attributed creator. Its provider field reads "Stealth." Its maker is silent. And in five days it processed over 160 billion tokens of developer queries - more throughput than many officially launched frontier models see in their first month.
The research community has a theory about whose model this is. The theory is plausible. It's not proven.
TL;DR
- Developers say Hunter Alpha is DeepSeek V4 - quietly launched on OpenRouter ahead of an April official launch
- Three technical signals point at DeepSeek; one historical precedent points equally hard at Zhipu AI
- 160+ billion tokens processed in five days, at zero cost, with all prompts logged by an entity that won't name itself
The Claim
On March 11, a model appeared on OpenRouter under the identifier openrouter/hunter-alpha. No company name. No developer attribution. Provider field: "Stealth." The next day, a companion model called Healer Alpha appeared from the same anonymous account.
Hunter Alpha's specs are serious. One trillion parameters. A 1,048,576-token context window - exactly 1 million tokens. Function calling, reasoning mode, tool use, structured output, multimodal input (text and image). Output speed: roughly 48 tokens per second. Price: $0 per million tokens in or out, in both directions.
Healer Alpha is different in character - multimodal with vision and audio, a 256K context window, and about 93 tokens per second output speed. Where Hunter Alpha runs long and deep, Healer Alpha runs lighter and faster across modalities.
Hunter Alpha appeared on OpenRouter on March 11 under a "Stealth" provider tag with no company attribution - then processed 160 billion tokens in five days.
Source: unsplash.com
The community consensus arrived fast. The combination of 1T parameters, a 1M context window, and strong agent-framework performance maps exactly onto what had been reported about DeepSeek V4 before any official announcement. Chinese outlet Whale Lab reported on March 16 that DeepSeek V4 is tracking for an April 2026 launch. A stealth deployment three to four weeks ahead of that fits a controlled load-testing pattern.
Neither DeepSeek nor OpenRouter has commented.
The Evidence
The Case for DeepSeek
Three technical signals distinguish this theory from ordinary speculation.
The knowledge cutoff match. Hunter Alpha's training data ends at May 2025. DeepSeek's own public chatbot has the same cutoff. This isn't a round number or an obvious estimate - it's a specific data point that would be hard to match by coincidence or by deliberately mimicking a known model.
The chain-of-thought signature. When Hunter Alpha reasons through a problem, it opens with "Hmm, the user said..." - a stylistic pattern that researcher Exocija documented in a side-by-side comparison on X. DeepSeek V3.2 uses the same opener. AI engineer Daniel Dewhurst, quoted in the Reuters wire story carried by The Star and Malay Mail, called this "the strongest signal" because "reasoning style is hard to disguise and tends to reflect how a model was trained."
Spec alignment with reported V4 targets. Before Hunter Alpha appeared, Chinese tech media described DeepSeek V4's planned architecture as roughly 1 trillion parameters with a 1-million-token context window. Hunter Alpha is exactly that. The dual-model structure - Hunter Alpha as the text and agentic flagship, Healer Alpha as the lighter omni-modal companion - matches the V4 Lite multimodal leak from early March.
Three technical fingerprints point toward DeepSeek: matching training cutoff, identical chain-of-thought opener, and specs that align with pre-release V4 reports.
Source: unsplash.com
The free-with-data-logging access model also fits. The OpenRouter listing is explicit: "All prompts and completions for this model are logged by the provider and may be used to improve the model." Offering free access to log high-quality developer queries at scale before an official launch is exactly the strategy a lab would use to improve a model in the weeks before release.
The Case Against
Independent benchmark researcher Umur Ozkul ran his own analysis and reached a different conclusion. He told Reuters that Hunter Alpha "is likely not DeepSeek V4," citing "differences in token behavior and architectural patterns when compared with DeepSeek's existing systems." He acknowledged the speculation was understandable given the timing and capability profile.
The stronger counter-argument isn't technical at all. It's precedent.
The same anonymous OpenRouter provider account that listed Hunter Alpha and Healer Alpha previously listed a model called Pony Alpha. Five days after Pony Alpha appeared on OpenRouter, Zhipu AI confirmed it was GLM-5. If this is the same provider - and the same account strongly suggests it's - Hunter Alpha could be GLM-6 and Healer Alpha its multimodal counterpart. Not DeepSeek.
This is a confirmed fact from recent history, not inference. It's the one data point the DeepSeek theory can't cleanly explain.
| Signal | What the Data Shows |
|---|---|
| Parameters | 1T - confirmed by OpenRouter listing |
| Context window | 1,048,576 tokens - confirmed |
| Self-identification | "Chinese AI model primarily trained in Chinese" |
| Knowledge cutoff | May 2025 - matches DeepSeek's public chatbot |
| Chain-of-thought opener | "Hmm, the user said..." - matches DeepSeek V3.2 |
| Spec match to V4 reports | Consistent with pre-release leaks; unconfirmed by DeepSeek |
| Provider identity | Claimed by no one; denied by no one |
| Pony Alpha precedent | Same account - confirmed Zhipu AI GLM-5 |
What They Left Out
The May 2025 cutoff and the CoT signature have dominated the coverage. What's getting less attention: the entire analysis rests on three named sources - Nabil Haouam, Daniel Dewhurst, and Umur Ozkul - all quoted in the same Reuters wire story, syndicated across dozens of outlets with no additional independent sourcing.
No third-party organization has run Hunter Alpha through a standardized evaluation. There are no published MMLU scores, no SWE-bench numbers, no systematic coding comparisons against DeepSeek V3.2 or any other known model. The "1T parameter" figure comes from the OpenRouter listing, which the anonymous provider wrote themselves. The model's actual parameter count and architecture remain unverified.
The access model also deserves more scrutiny than it's received. A developer sending queries to Hunter Alpha is handing that data to an entity that won't identify itself - for a model that won't say who trained it. The prompts are being used to improve a model. Which model, by which lab, for what purpose: unknown.
160 billion tokens of real developer queries now belong to someone. Whether that's DeepSeek, Zhipu AI, or a third party completely, the data collection is very deliberate.
The May 2025 cutoff match and the "Hmm, the user said..." pattern are the two strongest signals in the DeepSeek direction. Both are specific and reproducible. But the Pony Alpha precedent - the same account, previously confirmed as Zhipu AI - is the one fact neither the community consensus nor the Reuters sources have answered. Until DeepSeek or OpenRouter speaks, the model running at 48 tokens per second under a Stealth tag could belong to either lab. The 160 billion tokens of queries it collected in the meantime don't care which one it is.
Sources:
