MOSS-Music Beats Gemini on Music AI and Now Runs Locally
A community GGUF port of OpenMOSS's 8B music understanding model lets a C++/GGML desktop app caption your own audio dataset locally.
- scragnog released a GGUF conversion of OpenMOSS's MOSS-Music-8B-Instruct for the HOT-Step CPP desktop app.
- The model does music understanding, not generation: captions, lyrics ASR, chord and structural analysis.
- Base model scores 80.38 average on 8 music QA benchmarks, beating Gemini-3.1-Pro and Qwen3-Omni.
- Audio tower is kept f16; language model ships in q8_0 or f16 variants, Apache 2.0 licensed.
- Parity against fp32 PyTorch reference exceeds 0.9999 correlation at every module boundary.
- Tempo and key outputs are unstable; HOT-Step overrides them with Essentia DSP analysis.
MOSS-Music gets a local GGUF port
Developer scragnog has converted MOSS-Music-8B-Instruct to GGUF for local inference through HOT-Step CPP. The desktop application runs the model with a native C++/GGML engine, without a Python runtime, cloud API, or API key. GGUF is a model format designed for efficient inference and quantized weights, which reduces storage and memory requirements.
MOSS-Music is an open-source music understanding model from MOSI.AI, the OpenMOSS team, and Shanghai Innovation Institute. Its continual pretraining and supervised fine-tuning cover music captioning, lyrics transcription, structural analysis, chord and key detection, tempo reasoning, and long-form question answering. Its output is text, with audio generation outside the model’s scope.
Two files, one custom runtime
The conversion separates the model into a Qwen3-based language model and a mandatory f16 audio tower. HOT-Step’s custom engine connects the audio encoder, SwiGLU adapter, DeepStack merger, and language model. Stock llama.cpp builds do not implement that multimodal wiring, so the files target HOT-Step’s Model Manager.
| Component | Available formats | Purpose | Deployment note |
|---|---|---|---|
| Language model | q8_0, f16 |
Reasoning and text generation | q8_0 is the recommended default and is roughly half the size of f16. |
| Audio tower | f16 |
Audio encoding, feature adaptation, and layer injection |
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
