No.01 · 2026.10.15 · Unofficial field guide
Muse Spark OpenRunning it, sizing it, licensing it
Meta’s flagship open-weight model. How to run it, what hardware it needs, and what the license allows, with every number checked against the official model card.
Specifications
meta-models- Total parameters816B · MoE
- Active per token42B
- Experts8 of 256
- Context1M tokens
- Inputstext · image · video · audio
- Official weightsBF16 · NVFP4
- VRAM for NVFP4≥ 700 GB
- LicenseMuse Model License
Contents
5 guides- 01 OverviewArchitecture, modalities, inference settings, downloads
- 02 Run locallyMLX, vLLM, SGLang or llama.cpp, by hardware
- 03 License explainedThe $20M threshold, attribution and output ownership
- 04 Compare & chooseHow it differs from Muse Glimmer, and which to pick
- 05 Hosted APIsHosted APIs and clouds when you don’t self-host
How much hardware is enough?
4-bit · 32KEstimates at 4-bit with a 32K context. Try other setups in the calculator.
| Setup | Needed / usable | Verdict | Speed |
|---|---|---|---|
| 2 × Mac Studio M5 Ultra 512GB | 530 / 922 GB | ● Smooth | ~22 tokens/s |
| 8 × NVIDIA H200 141GB | 554 / 1072 GB | ● Smooth | ~217 tokens/s |
| 1 × Mac Studio M5 Ultra 512GB | 526 / 461 GB | ○ Not enough memory | — |
| 2 × RTX PRO 6000 + 512GB RAM | 530 / 182 GB | ◐ Runs, slowly | ~2.7 tokens/s |
One command to serve it
vLLMOn a multi-GPU server, the official vLLM command gives you an OpenAI-compatible endpoint.
vllm serve "meta-models/Muse-Spark-Open-816B-A42B" \
--tensor-parallel-size 8