Kunlun WanweiAnnounced today,Heavenly Craftsmanship 3.0 Large ModelThe performance has been significantly improved, and its Tiangong SkyMusic Musical ModelIt is also open to the public for testing today.
Tiangong 3.0 has 400 billion parameters, surpassing Grok-1 with 314 billion parameters.The world's largest open source MoE model. Tiangong 3.0 has significantly improved its performance in the fields of semantic understanding, logical reasoning, versatility, generalization, uncertainty knowledge, and learning ability, and its mathematical/reasoning/coding/cultural and creative capabilities have been improved by more than 30%. Tiangong 3.0 has added multiple AI capabilities such as multi-round search and comprehensive tool calls, chart drawing, research mode, enhancement mode, and image modification and expansion.
▲ Tiangong 3.0 model parameters surpass Grok-1
The Tiangong SkyMusic music model under Tiangong 3.0 is also open to the public today. Kunlun Wanwei said that Tiangong SkyMusic is "significantly" ahead of its competitors in the fields of vocal & BGM sound quality, vocal naturalness, pronunciation intelligibility, etc.Overall performance exceeds Suno V3, achieving the SOTA (State of the art model) for large music models, that is, the best performing model in current research.
SkyMusic adopts the Sora model architecture in the field of music and audio. Large-scale Transformer is responsible for composing music to learn the contextual dependencies of Music Patches and complete the controllability of music. Diffusion Transformer is responsible for singing. Through LDM, Music Patches are restored to high-quality audio, enabling it to support generation. 80 seconds 44100Hz sampling rate two-channel stereo song.
▲ SkyMusic AI music big model technical architecture
It is reported that SkyMusic has the following features:
High-quality AI music: Generate 80-second 44100Hz sampling rate two-channel stereo AI song
The human voice is "indistinguishable from the real thing": the Chinese level is extremely good, and the pronunciation is clear without any strange sounds
Lyrics paragraph control: The generated song can clearly distinguish the emotional changes in different lyrics paragraphs
Various music styles: support rap / folk / funk / ancient style / electronic, etc.
Intelligent music expression: Able to learn various singing techniques such as vibrato, opera, chanting, duet, automatic harmony, etc.
Reference music generation: Users upload their own reference music to generate songs with similar styles and singing styles
Dialect song generation: supports Cantonese, Chengdu dialect, Beijing dialect and many other dialects
According to public information, Kunlun Wanwei is a Chinese Internet platform company that has been developing overseas markets for more than ten years. Its business covers multiple fields including information distribution, social networking, entertainment, metaverse, games and AIGC. It has three major business segments, including AGI and AIGC, overseas information distribution and metaverse, and investment. Its markets cover China, Southeast Asia, Africa, the Middle East, North America, South America, Europe, etc. As of now, the average monthly active users worldwide are nearly 400 million, and overseas revenue accounts for 84%.