Skip to content
SPECIFICTION
Get listed

Khala (now Shao)

nanhuantech.com

WebAPISelf-hosted
Visit nanhuantech.com
Spec score 7.1 #115 of 123 in AI music generators

Spec sheet

Purpose
Shao, formerly Khala, is an open-source system for generating complete music from text descriptions and lyric conditions.Read 12 Oct 2026 · github.com
Intended users
The current release is mainly intended for researchers and developers familiar with GPU servers.Read 12 Oct 2026 · github.com
Audience
The current release is mainly intended for researchers and developers familiar with GPU servers and basic Docker and CUDA workflows.Read 12 Oct 2026 · github.com
Supported platforms
webRead 22 Sep 2026 · nanhuantech.com
Deployment
The README provides Docker setup instructions and a local web UI at 127.0.0.1:30869.Read 12 Oct 2026 · github.com
Support
The README invites users to join a WeChat group for discussion, usage questions, and future updates.Read 12 Oct 2026 · github.com
Commercial use
NoRead 22 Sep 2026 · nanhuantech.com
Compatibility limitation
Generation quality can vary with GPU architecture and software versions, and the project says bit-identical output across GPU types is not guaranteed.Read 12 Oct 2026 · github.com
Former name
The project was formerly called Khala and has been officially renamed Shao.Read 12 Oct 2026 · github.com
Full-song generation
It is designed to generate full-length musical works rather than short clips or loop-style accompaniment.Read 12 Oct 2026 · github.com
Generation architecture
Its pipeline uses a 64-layer residual vector quantization acoustic-token hierarchy, followed by a super-resolution model and decoder.Read 12 Oct 2026 · github.com
Generation controls
Natural-language prompts and lyrics can control style, mood, vocals, and content.Read 12 Oct 2026 · github.com
Generation pipeline
Its pipeline generates coarse acoustic tokens, completes higher RVQ token layers with a super-resolution model, and decodes the waveform.Read 12 Oct 2026 · github.com
GPU requirement
The full inference pipeline recommends an NVIDIA GPU with at least 24 GB of VRAM, such as an RTX 4090 or higher-tier GPU.Read 12 Oct 2026 · github.com
GPU requirements
The full inference pipeline recommends an NVIDIA GPU with at least 24 GB of VRAM, such as an RTX 4090 or higher-tier GPU.Read 12 Oct 2026 · github.com
Included system
The project includes a frontend UI, FastAPI backend dispatcher, single-GPU inference worker, model loading, and an end-to-end audio-generation path.Read 12 Oct 2026 · github.com
Model license
The linked model weights are listed as CC BY-NC 4.0.Read 12 Oct 2026 · huggingface.co
Model weights license
The model weights are stated to be intended for release under CC BY-NC 4.0.Read 12 Oct 2026 · github.com
Name change
The project was formerly called Khala and has been officially renamed to Shao, with its model, codebase, and research direction unchanged.Read 12 Oct 2026 · github.com
Prompt controls
Natural-language prompts and lyrics can control style, mood, vocals, and content.Read 12 Oct 2026 · github.com
Quality limitation
The project says inference quality is sensitive to GPU architecture, CUDA stack, container image, and Megatron/TransformerEngine versions, and bit-identical output across GPU types is not guaranteed.Read 12 Oct 2026 · github.com
Runtime dependencies
The listed runtime requirements include Docker, NVIDIA Container Toolkit, and a CUDA-compatible NVIDIA driver.Read 12 Oct 2026 · github.com
System components
The system includes a frontend UI, FastAPI backend dispatcher, and single-GPU inference worker.Read 12 Oct 2026 · github.com
Verified environments
The released pipeline is reported to produce normal results on an H800 with the reference image and to work on an RTX 4090 when the provided environment is followed.Read 12 Oct 2026 · github.com
Vocal generation
YesRead 22 Sep 2026 · nanhuantech.com

Where Specifiction places it

Alternatives to Khala (now Shao)All alternatives →