ModelBest unveils VoxCPM at China-ASEAN Expo: 2B speech model covering 30 languages

面壁智能 · wechat · 2026-09-28

At the 2026 China-ASEAN Expo, ModelBest (Mianbi) pitched its edge AI strategy built on its "density law" — model intelligence density doubling roughly every 3.5 months — with edge models already deployed across cars, phones, embodied AI and PCs, running on chips costing tens of yuan.

Its multilingual speech foundation VoxCPM uses a tokenizer-free, diffusion-autoregressive architecture: 2B parameters trained on 2.36M hours of multilingual data, supporting 30 languages and 9 dialects including Vietnamese, Thai, Malay, Lao and Burmese, with voice cloning, voice design, emotional expression, streaming output, LoRA fine-tuning and ComfyUI support. Target uses span multilingual call centers, real-time dubbing for Chinese firms going overseas, and rural broadcast. ModelBest also delivered Laos' national LLM under a "R&D in Beijing/Shanghai/Guangzhou, integration in Guangxi, deployment in ASEAN" path.

Original post →

More from Companies & People

Companies & People channel →