- Home
- DeepSeek News
- OpenRouter suddenly reveals stealth model blind test: Hunter & Healer twin stars suspected to be a leaked DeepSeek V4!

OpenRouter suddenly reveals stealth model blind test: Hunter & Healer twin stars suspected to be a leaked DeepSeek V4!
OpenRouter quietly launched two stealth models, Hunter Alpha and Healer Alpha, with 1T parameters, 1M context, and 256k lightweight full-modality features, directly targeting the flagship and Lite versions of DeepSeek V4.
The world of AI hums with anticipation as OpenRouter, a leading AI model aggregation platform, stealthily launches two mysterious models in "Stealth Mode": Hunter Alpha and Healer Alpha.
These models, with names steeped in geek romanticism, boast performance parameters that have sent shockwaves through the AI community. The burning question on everyone's minds: Could these be the highly anticipated blind test versions of DeepSeek V4?!
đĩī¸ââī¸ Deep Dive: Specification Alignment and the Chain of Evidence
Editor's investigation, through cross-referencing official OpenRouter descriptions with developer test data, has unearthed three decisive pieces of "irrefutable" evidence:
1. The "Absolute Match" of 1 Trillion Parameters and 1M Context
OpenRouter's description of Hunter Alpha clearly states: 1 Trillion (1ä¸äēŋ) Parameters + 1M Token Context. This aligns perfectly with the core parameters of DeepSeek V4 (1T MoE architecture, 1M Context) leaked on HuggingFace yesterday. Currently, no other unreleased model on the market that simultaneously meets both criteria exists except DeepSeek V4.
2. A Dual-Track Strategy: "Flagship" and "Lightweight?"
In addition to specification alignment, early community discussions suggested a potential "flagship + lightweight" combination test: Hunter Alpha corresponding to the fully-fledged flagship version with 1 million context tokens, while Healer Alpha corresponds to the lightweight, all-modal version with 256,000 context tokens.
đ Community Rashomon: Who Exactly IS Healer Alpha?
As the blind test intensifies, the AI community is embroiled in a three-way "Rashomon" conjecture about the identity of its twin, Healer Alpha:
- Faction A: DeepSeek V4 Lite (SeaLion-Lite) This is the most logical inference. The 256,000 context window aligns closely with the previously leaked Lite version specifications, focusing on all-modal interaction and forming a high-low pairing with the flagship version.
- Faction B: Xiaomi MiMo V3 (Xiaomi's Counterattack) Some developers point out that Healer's exceptionally strong "Action" capabilities are very similar to Xiaomi's latest MiMo (Mobile Intelligence, Multimodal Optimization) architecture. Given Xiaomi's recent showcase of its whole-house intelligent AI at MWC 2026, this model could very well be a test "sleeper agent" deployed by Xiaomi to showcase its AGI hardware ecosystem.
- Faction C: Moonshot Kimi K3 (Moonshot AI's Retaliation) Kimi K3's release window is in March. Healer Alpha's demonstrated strict adherence to instructions and logical stability bear a resemblance to the Kimi family's hallmark traits.
3. Explicit Support for Agent Frameworks Like OpenClaw
Most exciting for developers, the description of Hunter Alpha explicitly mentions "Reliability and instruction-following precision required for frameworks like OpenClaw." As one of the most popular Agentic frameworks currently available, OpenClaw demands exceptional rigor in model logic chains. DeepSeek has clearly prioritized "Agent Productivity" during the fine-tuning phase of V4.
đŽ Geek Romance Behind the Gamified Naming
Those familiar with the DeepSeek team know that this Chinese team is deeply influenced by geek culture and anime/game culture.
- Hunter: Represents proactive action and precise execution, corresponding to the "Action" capabilities of an Agent.
- Healer: Represents perceiving everything and providing all-around support, corresponding to the "Understanding" capabilities of a full-modal model.
This MMORPG-colored naming scheme not only cleverly conceals their true identities but also fills this blind test with an aura of mystery.
đ How to Get Early Access?
Currently, both models are in Free blind test phase on OpenRouter. If you want to experience the intellect of V4 ahead of the official launch, now is the best time.
- Developer Channel: Head directly to OpenRouter and search for
stealth/hunter-alphato make the call. - Stable Alternative: If OpenRouter blind test traffic is overloaded, consider following AIML API. As a long-standing partner of DeepSeek, AIML API often offers the most stable official interface in the first instance.
- Efficient Tools: Users of the Monica Plugin can pay attention to its model list updates; Monica has a habit of quickly integrating OpenRouter's high-quality new models.
đĄ Editorial Commentary
With the launch of the Stealth models, the official release countdown for DeepSeek V4 has entered "hour-level". All indications point towards V4 being more than just a smarter chatbot; it is the ultimate "Hunter" built for the Agent era.
Follow DeepSeekV4.app, Editor will monitor every global traffic surge for you!
Author

Table of Contents
More Posts

DeepSeek V4 Release Date Set? Unveiling Huawei Ascend 950PR Computing Base and 1T Parameter Specs - The Final Puzzle Piece
The trending DeepSeek V4 release date might soon be revealed. Deep dive into how V4 utilizes Huawei Ascend 950PR chips to achieve computing decoupling, with early disclosure of 1T parameters, 1M context, and other core specs and benchmarks.


Breaking Leak: DeepSeek V4 Full Specifications Revealed, 1 Trillion Parameter MoE Monster Arrives!
After upload records appeared on HuggingFace, DeepSeek V4's core specifications leaked: 1 trillion parameters, 1 million context window, native audio support.

![[Deep Prediction] When will DeepSeek V4 be released? Sniping GPT-5.4, three clues point to tomorrow (Tuesday). [Deep Prediction] When will DeepSeek V4 be released? Sniping GPT-5.4, three clues point to tomorrow (Tuesday).](/_next/image?url=%2Fimages%2Fnews%2F2026-03-09-deepseek-v4-release-date-prediction.png&w=3840&q=75&dpl=dpl_DTeDfmyHMdFGtyRv2TK8B1Vq3xKY)
[Deep Prediction] When will DeepSeek V4 be released? Sniping GPT-5.4, three clues point to tomorrow (Tuesday).
14 months have passed since the release of DeepSeek-R1. Facing the strong attack from OpenAI's GPT-5.4, DeepSeek V4 is on the verge of release. This article dissects three key clues to reveal why the trillion-parameter domestic large language model is highly likely to launch on March 10th.

Newsletter
Join the community
Subscribe to our newsletter for the latest news and updates