The mysterious AI model 'Hunter Alpha,' with its trillion-parameter scale, became a hot topic, leading to speculation that DeepSeek was secretly testing 'V4,' but it was later revealed to belong to Xiaomi.
It has been discovered that a large-scale AI model with 1 trillion parameters, ' Hunter Alpha, ' has been newly registered on 'OpenRouter,' a platform that provides access to various AI models. The developer is completely unknown, which led to speculation that it might be from China's DeepSeek, but its true identity was later revealed.
A mystery AI model has developers buzzing: Is this DeepSeek's latest blockbuster? | Reuters
Hunter Alpha - API Pricing & Providers | OpenRouter
'Hunter Alpha' has no attribution to the developer, and was initially introduced as a 'stealth model' on OpenRouter.
When Reuters used this model, it described itself as a 'Chinese AI model primarily trained in Chinese' and stated that it was trained with data up to May 2025. This period is the same as other DeepSeek models, and considering that it is trained in Chinese, Reuters speculated that it might be one of DeepSeek's models.
The Hunter Alpha introduction page explains that the model has a scale of 1 trillion parameters. This means the system was trained using approximately 1 trillion tunable values, and models with a large number of parameters generally require considerable computational power to operate. The model is also said to have a context window of up to 1 million tokens.
Reuters reported on March 18, 2026, that 'these specifications are similar to those of DeepSeek's next-generation model, the 'V4,' which has already been reported and is expected to be unveiled as early as April.'
However, on the same day as the Reuters report, information on Hunter Alpha was updated, revealing that it was actually an early test version of the ' MiMo-V2-Pro ' developed by Xiaomi.
OpenRouter Xiaomi: MiMo-V2-Pro
https://openrouter.ai/xiaomi/mimo-v2-pro
Details about the MiMo-V2-Pro are available on the following page.
MiMo-V2-Pro | Xiaomi
https://mimo.xiaomi.com/mimo-v2-pro
The MiMo-V2-Pro boasts over one trillion parameters and a context length of one million, making it a highly optimized model for agent scenarios. It is highly adaptable to general-purpose agent frameworks such as OpenClaw, and boasts world-class performance in agent benchmarks such as PinchBench and ClawBench, with perceived performance comparable to Anthropic's Claude Opus 4.6.
Artificial Analysis, a website that compares the performance of AI models, has also published
Xiaomi stated, 'Its coding capabilities surpass those of Claude 4.6 Sonnet, and the overall agent performance is approaching that of Claude Opus 4.6. The stability and accuracy of tool calls have also been significantly improved. We released it to OpenRouter a week ago under the codename 'Hunter Alpha,' and the number of API calls has steadily increased, maintaining the top spot on the daily chart for several days. Hunter Alpha is an initial internal test build of MiMo-V2-Pro. As a result of continuous iteration and optimization over the past week, MiMo-V2-Pro has shown significant improvements in long-context functionality and agent scenario stability.'
Related Posts:
in AI, Posted by log1p_kr