Muse Spark Open 816B — VRAM, speed & local setup
Written by Jakub Rusinowski · Last updated
816B parameters (822B in the published tensors, counting the 1.8B perception encoder and untied embeddings) with 42B active per token: 256 experts, 8 routed per token, 62 layers with grouped-query attention (16 KV heads of dimension 64). 1,048,576-token context. Meta's card names SGLang, vLLM, Transformers, TensorRT-LLM, Unsloth and llama.cpp, and says it is optimised to run on two Mac Studios with M5 Ultra under MLX. The repository is gated: Meta asks for your full legal name, date of birth and organisation before granting access.
Muse Spark Open 816B needs about 497 GB of VRAM at Q4_K_M — quantized weights plus framework overhead, before any KV cache. On Apple Silicon that figure comes out of unified memory.
VRAM and speed by quantization
Quoted against NVIDIA RTX 4090 (24 GB). Includes the KV cache at 8K context, so it reads higher than the headline figure.
| Quant | Memory | VRAM | Speed (est.) | Fit |
|---|---|---|---|---|
| Q2_K 2.63 bpw | 273.1 GB | — | Too big | |
| Q3_K_M 3.41 bpw | 353.3 GB | — | Too big | |
| Q4_K_M 4.83 bpw | 499.2 GB | — | Too big | |
| Q5_K_M 5.67 bpw | 585.5 GB | — | Too big | |
| Q6_K 6.56 bpw | 677 GB | — | Too big | |
| Q8_0 8.50 bpw | 876.3 GB | — | Too big | |
| F16 16.00 bpw | 1647 GB | — | Too big |
Black marker = usable memory on the NVIDIA RTX 4090 (24 GB). Estimates from the memory-bandwidth roofline on the methodology page. Muse Spark Open 816B VRAM calculator →
Get Muse Spark Open 816B running
Muse Spark Open 816B is not practical on a single consumer GPU: it needs about 497.1 GB at Q4_K_M, more than any single device in our catalogue holds with headroom. How multi-GPU inference works, or rent a multi-GPU server by the hour:
As an Amazon Associate we earn from qualifying purchases. Cloud GPU links are referral links — we may earn a commission at no extra cost to you.
How to run Muse Spark Open 816B
No local Ollama tag. Ollama lists muse-spark-open only as a cloud-hosted ':cloud' tag, which runs on Ollama's servers, not your machine.
Specifications
Verified — Checked against the primary source — the model card or the vendor spec page — and corroborated by a second independent source.
- Parameters
- 816 Billion (42B active)
- Context window
- 1,048,576
- Architecture
- Mixture-of-Experts (256 experts, 8 active), GQA, perception encoder
- Provider
- Meta
- Licence
- Muse Model License
- Availability
- Weights downloadable
- Access
- Gated: request access on Hugging Face and accept the licence first
- Practical deployment
- Multi-GPU server
- Specified at
- Q4_K_M
- System RAM
- 1024 GB
- Record updated
- 2026-10-15
Weights are downloadable and commercial use is permitted, subject to the licence’s acceptable-use terms.
Model-as-a-service operators above $20M gross revenue in any 12 months must notify Meta and sign a separate agreement; products must display "Built with Muse Spark Open".
Limits and caveats
- Gated repository: access is granted per request, and the form asks for legal name, date of birth and organisation.
- Muse Model License, not an OSI licence: products must display "Built with Muse Spark Open", and model-as-a-service operators above $20M revenue need a separate agreement with Meta.
- Not practical on any single consumer device: Meta's own minimum is 700 GB of GPU memory for the NVFP4 checkpoint and about 2 TB for BF16.
- The NVFP4 checkpoint needs Blackwell-class GPUs under vLLM.
See also Muse Glimmer 30B
Quality and use cases
Scores as published by the model’s authors or an independent evaluator — quality, not throughput, and not measured by us.
Can I run Muse Spark Open 816B on my GPU?
Muse Spark Open 816B — frequently asked questions
How much VRAM does Muse Spark Open 816B need?
About 497 GB at Q4_K_M — quantized weights plus framework overhead, before any KV cache. The cache grows with context length and is added on top; the table above folds it in. Apple Silicon counts unified memory toward the same figure.
Does Muse Spark Open 816B run on an RTX 4090 (24 GB)?
No. Muse Spark Open 816B needs about 497 GB at Q4_K_M, more than a single RTX 4090's 24 GB. It needs a larger card, several GPUs, or Apple Silicon with enough unified memory — or it runs with part of the weights offloaded to system RAM, which is much slower.
How do I run Muse Spark Open 816B locally?
No local Ollama tag. Ollama lists muse-spark-open only as a cloud-hosted ':cloud' tag, which runs on Ollama's servers, not your machine. Running the published tag would send your prompts to a hosted GPU rather than your own machine.