Speed t/s31.9
Total$40,578
- RAM
- 384 GB DDR5-480012 × 32 GB · 461 GB/s
- $15,348
- Live price
- Host
- AMD EPYC 9354Supermicro H13SSL-N motherboard
- $2,700
- Reference price
- Power
- 1,200 WAbout 980 W for the whole system
- $180
- Reference price
How to Deploy Muse Spark Open Locally
Muse Spark Open, released by Meta in October 2026, is the open-weight version of Muse Spark: an 816B MoE that uses 42B parameters per token. It reads images, video and audio, and 4-bit with 32K of context is about 509 GB, so at home it takes eight 96GB GPUs, a 512GB Mac Studio or a server with lots of RAM. llama.cpp support hasn't reached mainline yet, so Unsloth's GGUFs currently need a build from the patch branch.
Muse Spark Open details
- Min VRAM at 4-bit508 GB · 8K
- Parameters816B
- Active params42B
- Max context1M
- Architecture
- MoE
- Experts offloaded
- 14 GB VRAM + 495 GB RAM
- License
- Muse Model License
- Launch date
- Oct 15, 2026
- Vendor
- Meta AI
- Hugging Face repo
- meta-models/Muse-Spark-Open-816B-A42B
- NVFP4 repo (Blackwell)
- meta-models/Muse-Spark-Open-816B-A42B-NVFP4
- Layers
- 62
- Hidden size
- 8,192
- KV heads
- 16
- Head dim
- 64
How much VRAM does Muse Spark Open need for local deployment?
4-bit with 32K of context needs about 509 GB, and 2-bit brings that down to about 315 GB. Longer context barely adds to it: three out of four layers only remember the last 2,048 tokens, and just one layer in every four looks at the whole context, so taking 4-bit from 32K all the way to 1M adds only about 67 GB.
| Quantization | WeightsGB | 32K contextGB | 128K contextGB | 256K contextGB | 1M context (max)GB |
|---|---|---|---|---|---|
| 1-bit | 208.9GGUF | 212.5 | 218.9 | 227.5 | 279.1 |
| 2-bit | 311.2GGUF | 314.8 | 321.2 | 329.8 | 381.4 |
| 3-bit | 401.5GGUF | 405.1 | 411.5 | 420.1 | 471.7 |
| 4-bit | 505.5GGUF | 509.1 | 515.5 | 524.1 | 575.7 |
| 5-bit | 591.2est. | 594.8 | 601.2 | 609.8 | 661.4 |
| 8-bit | 871.7GGUF | 875.3 | 881.7 | 890.3 | 941.9 |
| BF16/FP16 | 1,640.3GGUF | 1,643.9 | 1,650.3 | 1,658.9 | 1,710.5 |
Full calculation rules are on the Methodology page.
Can a single GPU run Muse Spark Open?
Not on any single graphics card; the only machine that holds it is the 512GB Mac Studio. Even Unsloth's smallest 1-bit build is about 176 GB, more than any graphics card and more than any 128GB machine. A single 512GB Mac Studio M3 Ultra has the capacity for 4-bit, as long as you raise the share of memory macOS lets the GPU use; we estimate about 21 tokens/s. Meta's own recommendation is two 512GB Mac Studio M5 Ultras linked together with MLX.
| Device | Price | Speed t/s | Max context |
|---|---|---|---|
| Mac Studio M3 Ultra 512GBWhole system · 512 GB | $22,85012 listings | 32.3 | 1M |
| Device | Price | Speed t/s | Max context |
|---|---|---|---|
| Mac Studio M3 Ultra 512GBWhole system · 512 GB | $22,85012 listings | 21.3 | 637K |
Only common GPUs and systems are listed, one set for models under 50B and another for larger ones, sorted by price from low to high.
How many GPUs does Muse Spark Open need?
The entry point for 2-bit is four 96GB RTX PRO 6000s at about 58 tokens/s; 4-bit needs eight of them and runs at about 41, and eight 80GB A100s work too. Two 512GB Mac Studio M3 Ultras linked over Thunderbolt 5 hold both 4-bit and 8-bit, at about 21 and 13 tokens/s, the same approach as Meta's recommended pair of Mac Studios. The official vLLM and SGLang recipes run Meta's own 4-bit release, about 530 GB, and are written for data center hardware such as four Blackwell Ultra GPUs or eight H200s.
| Device | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|
 DGX Spark 128GB128 GB per unit | 4 units | $34,888 | 512 GB | 11.5 | 960 W |
| Instinct MI21064 GB per card | 8 cards | $39,976 | 512 GB | 54.2 | 2,400 W |
| RTX A600048 GB per card | 8 cards | $40,000 | 384 GB | 29.4 | 2,400 W |
| RTX PRO 5000 72GB72 GB per card | 8 cards | $85,776 | 576 GB | 46.6 | 2,400 W |
| RTX PRO 600096 GB per card | 4 cards | $89,400 | 384 GB | 57.9 | 2,400 W |
| A100 80GB PCIe80 GB per card | 8 cards | $143,992 | 640 GB | 61.2 | 2,400 W |
| GeForce RTX 4090 48GB48 GB per card | 8 cards | No price | 384 GB | 37 | 3,600 W |
| Device | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|
| Mac Studio M3 Ultra 512GB512 GB per unit | 2 units | $45,700 | 1,024 GB | 20.8 | 960 W |
| RTX PRO 5000 72GB72 GB per card | 8 cards | $85,776 | 576 GB | 32.1 | 2,400 W |
| A100 80GB PCIe80 GB per card | 8 cards | $143,992 | 640 GB | 43.3 | 2,400 W |
| RTX PRO 600096 GB per card | 8 cards | $178,800 | 768 GB | 40.7 | 4,800 W |
| Device | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|
| Mac Studio M3 Ultra 512GB512 GB per unit | 2 units | $45,700 | 1,024 GB | 12.8 | 960 W |
Mac Studios link up over Thunderbolt 5 and run with MLX or exo; vLLM doesn't support them. Full calculation rules are on the Methodology page.
How do you run Muse Spark Open using system RAM?
With the experts in system RAM, the GPU only has to keep a dozen or so GB: the 256 experts make up 98% of the weights, and there's no always-on shared expert. The hard part is having enough RAM. Only one configuration in the table can hold it, a 12-channel DDR5 server with one RTX PRO 6000: 2-bit with 384GB of RAM runs at about 32 tokens/s, while 4-bit needs all 768GB and drops to about 19.
Speed t/s19.1
Total$55,230
- RAM
- 768 GB DDR5-480012 × 64 GB · 461 GB/s
- $30,000
- Live price
- Host
- AMD EPYC 9354Supermicro H13SSL-N motherboard
- $2,700
- Reference price
- Power
- 1,200 WAbout 980 W for the whole system
- $180
- Reference price
Speed t/s10.8
Total$55,230
- RAM
- 768 GB DDR5-480012 × 64 GB · 461 GB/s
- $30,000
- Live price
- Host
- AMD EPYC 9354Supermicro H13SSL-N motherboard
- $2,700
- Reference price
- Power
- 1,200 WAbout 980 W for the whole system
- $180
- Reference price
As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.
What to know before running Muse Spark Open locally
llama.cpp and vLLM support are still waiting to merge
llama.cpp support was only submitted on release day (PR #30616) and hasn't been merged into mainline, so you have to build that branch before Unsloth's GGUFs will load, and Ollama has no official tag yet. The patch author tested text, images, audio and tool calling on an M3 Ultra and an RTX PRO 6000. Images and audio need separate projector files; Unsloth's repository only has the text part for now, but the patch author's GGUFs on ggml-org include them. Add --jinja for chat. vLLM support isn't in a release yet either (PR #61879): the official recipe says to use vLLM's muse-spark-open Docker image for now, and the validated setup runs Meta's 4-bit release on eight GPUs.
Unsloth's quants
Unsloth has published eight GGUFs from 1-bit to 4-bit; the 4-bit UD-Q4_K_XL is about 506 GB. The 8-bit and BF16 rows come from the ggml-org files; nobody has made a 5-bit build yet, so that row is estimated from the parameter count. Unsloth says the 2-bit UD-IQ2_XXS (about 226 GB) runs on a machine with 256GB of RAM, though with the experts squeezed to around 2 bits, answer quality takes a hit.
It thinks the longest by default
Set thinking depth with reasoning_effort: low, medium, high or xhigh. The default is the maximum, xhigh. Meta recommends high or xhigh for coding and agent work and low or medium when you want faster replies; any other value returns an error. The recommended sampling is temperature 1.0, top_p 1.0 and no top_k limit. Unsloth's GGUFs ship with a default top_k of 64, so add --top-k 0 to match Meta's settings.
Not the Llama license
It uses Meta's new Muse Model License. You can use it commercially, fine-tune it and distill it, but products must prominently show “Built with Muse Spark Open.” A company that offers model API access to third parties and whose group revenue tops 20 million US dollars over any 12 months has to sign a separate agreement with Meta first; internal use and powering features in your own product aren't covered by that clause. The official Hugging Face repository requires manual approval before you can download, while Unsloth's GGUFs don't.
Launch commands
Ollama
The Ollama library has no official tag for it yet.
llama.cpp
Follow the deployment guide to prepare the required runtime and hardware.
llama-server -hf unsloth/Muse-Spark-Open-816B-A42B-GGUF:UD-Q4_K_XL --jinja -c 8192
vLLM
Follow the deployment guide to prepare the required runtime and hardware.
docker run --gpus all \
-p 8000:8000 \
--ipc=host \
--env "HF_TOKEN=$HF_TOKEN" \
-v ~/.cache/huggingface:/root/.cache/huggingface \
vllm/vllm-openai:muse-spark-open meta-models/Muse-Spark-Open-816B-A42B-NVFP4 \
--tensor-parallel-size 8 \
--enable-expert-parallel \
--max-model-len 1048576 \
--reasoning-parser muse \
--tool-call-parser muse \
--enable-auto-tool-choiceFile names follow that week's model-library snapshot; the definitions are in the methodology. Methodology
Other models in the Meta family
| Model | Total params | Active params | 4-bit size |
|---|---|---|---|
| Llama 3.2 3BDense | 3.2B | 3.2B | 2 GB |
| Llama 3.1 8BDense | 8B | 8B | 4.9 GB |
| Llama 3.3 70BDense | 70.6B | 70.6B | 42.5 GB |
| Llama 4 Scout 17B 16EMoE | 108.6B | 17B | 65.4 GB |