Cai Haoyu's AI company Anuttacon releases its first video large model LPM1.0

ME News message. On April 11 (UTC+8), the AI company Anuttacon, founded by Mihoyo founder Cai Haoyu, quietly released a video character performance large model, LPM 1.0. It supports real-time full-duplex audio and video conversations, and can simultaneously speak, sing, listen, and respond with emotions. The mouth movements, microexpressions, and body rhythm details all achieve a high level of realism.

In terms of technology, LPM 1.0 uses a 17B-parameter Base model and distills it into a low-latency streaming Online version. It supports identity-consistent generation of any duration, and is compatible with a variety of character styles—including realistic, 2D anime, 3D games, and non-humanoid creatures—without the need for fine-tuning. The team also simultaneously released the character performance evaluation benchmark, LPM-Bench, achieving SOTA results across multiple metrics.

Compared with Kling-Avatar 2.0 and OmniHuman 1.5, which have a maximum generation limit of 30 seconds, LPM 1.0 supports unlimited-duration output and is positioned for scenarios such as dialogue agents, game NPCs, and virtual live streaming.

Notably, Anuttacon explicitly states that this release is only for academic exchange, and there are currently no plans to open source or provide APIs, demos, or any products externally. The general view is that one of the core purposes of this public release is to attract AI technical talent.

LPM 1.0 is Anuttacon’s latest move in the AI game engine and intelligent NPC track since the company’s founding, and it is also another phased achievement in Cai Haoyu’s ongoing deepening layout—from Mihoyo’s AI tool development to gradually expanding the underlying model capabilities of the AnuNeko chat product. (Source: BlockBeats)