Suno at home, now without the 7.6 GB of Python

What it does

It generates full songs with vocals — up to ten minutes — from a caption and lyrics, or just a one-line idea, entirely on your own NVIDIA GPU. Around the generator sits a whole studio: covers, remixes, stem separation, MIDI transcription, word-level karaoke, video clips, LoRA training, and a library with tags and cover art. It runs fully offline on Windows, from 4 GB of VRAM.

The interesting bit

Version 2.0 dropped the Python/Node runtime entirely: the engine is acestep.cpp (C++ on GGML), wrapped in a Rust service with a React UI in a Tauri window — one executable, no Gradio, no browser tab. That swap cut the minimum card from 12 GB to 4 GB via GGUF quantisations from Q4 (3.1 GB) up to BF16. There’s also an MCP server with 164 tools, so an agent like Claude can drive the entire studio, and a Winamp mode with MilkDrop — because apparently someone still remembers what a music player should feel like.

Key highlights

• Full songs with vocals in ~33–36 s for a ~3-minute track on an RTX 4090 (BF16, XL turbo), loading included
• Every ACE-Step 1.5 DiT — 2B and XL, turbo, SFT, merges, community fine-tunes — in Q4 to BF16, switchable in the UI
• Train your own LoRA on-card with a three-step wizard; lyrics fetched from LRCLIB and friends, songs described by MOSS-Music
• Exact replay: every track keeps its request and audio codes for bit-for-bit re-renders or new takes
• Six-stem separation (HT-Demucs), any track to multi-instrument MIDI, VST3 plugin chains, MP4 video clips

Caveats

• Windows-only build; macOS and Linux users are pointed back at the 1.x release or acestep.cpp itself
• Some 1.x features didn’t survive the rewrite: flow-edit, the “Most Natural” repaint mode, the model tools, and free Pollinations covers
• The feature list is enormous — expect a learning curve proportional to it

Verdict

If you have an NVIDIA card and want Suno-style songs without sending lyrics to anyone’s API, this is the most complete local package around. If you’re not on Windows or just want a quick demo, the underlying acestep.cpp is the leaner path.