MODEL · OPENMOSS / MOSI.AI · ~9B TOTAL (8B LLM + AUDIO ENCODER)
MOSS-Music-8B (Instruct + Thinking)
The first open-weight music-understanding LLM worth flagging — does lyrics ASR with time-aligned transcription, musical captioning, key/tempo/chord reasoning, structural analysis (intro/verse/chorus/bridge/outro), instrument + voice recognition, and music QA. Audio encoder runs at 12.5 Hz temporal resolution. 80.38% avg accuracy across 8 music-QA benchmarks; 15.88% avg WER/CER on lyrics; 4.36/5.0 MusicCaps captioning. Thinking variant adds chain-of-thought reasoning over audio.
License: Apache 2.0 · Context: inherits 8B backbone · Released: May 1, 2026
The decision in five lines
- The call
- Consider — runnable locally, family reference
- Best for
- Local evaluation and family reference
- Runs on
- 23 hardware picks fit (cheapest: Intel Arc B580 12 GB · $249)
- Watch out
- You want music GENERATION — this is understanding-only (audio in, text out).
- Evidence
- Estimated
- ~9B total (8B LLM + audio encoder)
- PARAMETERS
- MUSIC UNDERSTANDING
- TYPE
- inherits
- CONTEXT
- ~6–8 GB (Q4) / ~18 GB (FP16, est.)
- VRAM AT Q4
Where we recommend this
This model isn’t currently in an active planner slot. See the runner notes below if you’re running it anyway.
The call
The first open-weight music-understanding LLM worth flagging — does lyrics ASR with time-aligned transcription, musical captioning, key/tempo/chord reasoning, structural analysis (intro/verse/chorus/bridge/outro), instrument + voice recognition, and music QA. Audio encoder runs at 12.5 Hz temporal resolution. 80.38% avg accuracy across 8 music-QA benchmarks; 15.88% avg WER/CER on lyrics; 4.36/5.0 MusicCaps captioning. Thinking variant adds chain-of-thought reasoning over audio.
When not to use: You want music GENERATION — this is understanding-only (audio in, text out). For text-to-music try Suno / Udio (hosted) or Stable Audio (open-weight, separate model). And it's English-leaning on lyrics ASR; non-Latin script lyrics may suffer.
Runner notes
Recommended runtime is SGLang Serving (`sglang serve --model-path ./weights/MOSS-Music-8B-Instruct --trust-remote-code`). PyTorch 2.9+cu128 + FlashAttention 2 + FFmpeg 7. Python 3.12. Gradio app for local UI. No GGUF / Ollama route yet (non-standard audio-LLM arch). Sibling MOSS-Music-8B-Thinking for reasoning over audio.
Hardware that fits
Every hardware pick whose memory fits this model at the quant we recommend. Sorted cheapest-first — the top row is your best-value fit. Click through for the full buyer’s guide.
- Intel Arc B580 12 GBPerfect · 1.7× 12 GB · $249–$299
- NVIDIA RTX 3060 12 GBPerfect · 1.7× 12 GB · $250–$340
- Minisforum UM890 ProPerfect · 3.3× 32 GB DDR5 (shared) · $463–$580 all-in
- RTX 5060 Ti 16 GBPerfect · 2.2× 16 GB · $560–$610
- AMD Radeon RX 9070 XTPerfect · 2.2× 16 GB · $649–$849
- AMD Radeon RX 7900 XTXPerfect · 3.3× 24 GB · $770–$1,400
- Mac Mini M4 16 GBPerfect · 1.5× 16 GB unified · $799 (new floor) / $499–$599 (eBay/residuals)
- NVIDIA RTX 3090 (used, single)Perfect · 3.3× 24 GB · $800–$1,000
- NVIDIA RTX 5070 TiPerfect · 2.2× 16 GB · $980–$1,300
- NVIDIA RTX 5080Perfect · 2.2× 16 GB · $999–$1,250
- MacBook Air M5 24 GBPerfect · 2.2× 24 GB unified · $1,299–$1,699
- Mac Mini M4 Pro 24 GBPerfect · 2.2× 24 GB unified · $1,399
- Dual RTX 3090 (used)Perfect · 6.6× 48 GB · $1,800–$2,500 all-in
- Framework Desktop (Ryzen AI Max+ 395)Perfect · 11.9× 128 GB unified · $1,999–$2,851
- NVIDIA RTX 4090Perfect · 3.3× 24 GB · $2,200–$2,800
- M5 Pro MacBook Pro 48 GBPerfect · 4.5× 48 GB unified · $2,599–$3,099
- Mac Studio M4 Max 64 GBPerfect · 5.9× 64 GB unified · $3,199
- NVIDIA RTX A6000 (48 GB, used)Perfect · 6.6× 48 GB ECC · $3,500–$4,500
- NVIDIA RTX 5090Perfect · 4.4× 32 GB · $3,800–$4,100
- Mac Studio M3 Ultra 96 GBPerfect · 8.9× 96 GB unified · $3,999
- M5 Max MacBook Pro 64 GBPerfect · 5.9× 64 GB unified · $4,499
- NVIDIA DGX SparkPerfect · 11.9× 128 GB unified · $4,699
- Dual RTX 5090Perfect · 8.9× 64 GB (2×32) · $8,500–$10,500
Next step
Find-by-model — see what hardware runs this→