2026-10-16
Meta Muse Spark Open: 816B multimodal, even locally
Meta's new open-weight MoE reads text, images, audio and video. We examine the architecture and the hardware it actually needs, from a 512 GB Mac to four Blackwell GPUs.
Muse Spark Open is Meta's new frontier open-weight model for agentic coding and multimodal reasoning. It accepts text, images, audio and video, uses tools and generates text. It joins the SophyAI catalog with two reference hardware profiles.
What is inside
It is a Mixture of Experts with about 816 billion parameters, while activating around 42 billion per token. Each of its 62 layers contains 256 experts and the router selects 8. Compute per token is therefore far below the total, but every expert must still remain in memory.
| Feature | Value |
|---|---|
| Total / active parameters | ~816B / ~42B |
| Experts | 256 per layer, 8 active |
| Input | text, images, audio, video |
| Output | text |
| Native context | up to 1,048,576 tokens |
| Checkpoints | BF16 and NVFP4 |
Images and video frames pass through a 1.8B-parameter perception encoder. Audio becomes 12.5 tokens per second. The model is trained for long agentic workflows: reading repositories, editing multiple files, running tools, observing results and correcting itself.
Memory is the real constraint
The BF16 checkpoint occupies about 1.5 TB. Meta and Hugging Face document Transformers inference on 4 nodes with 8 H200 GPUs each, or 32 GPUs. A 1M context creates another constraint: its BF16 KV cache is about 250 KB per token and can reach roughly 266 GB.
NVFP4 reduces the footprint substantially but requires NVIDIA Blackwell. The reference vLLM configuration uses 4 Blackwell GPUs, tensor parallel 4 and expert parallel, with a 131,072-token serving context. It supports text and images.
For all modalities, SGLang is documented on 4×GB300. Hugging Face also offers prevalidated profiles using 8×RTX PRO 6000 for 128k or 8×H200 for the full 1M context.
Can it really run locally?
Yes, although “local” means a Mac far beyond most personal machines. The Q2_K GGUF is about 304 GB across 22 shards. Hugging Face reports:
- a Mac Studio M3 Ultra with 512 GB unified memory;
- about 304 GB used by the model;
- roughly 18 tokens/s with llama.cpp;
- text, image and audio input.
The MLX oQ3e quant uses about 344 GB and reaches roughly 21 generation tokens/s on the same Mac. A 256 GB Mac is not enough. Even with 512 GB, context must be constrained because weights and KV cache share the same memory.
SophyAI catalog profiles
We added two visible but not yet launchable recipes:
- vLLM NVFP4 on 4×B200/B300, 128k, text and images.
- llama.cpp Q2_K on a 512 GB M3 Ultra Mac, text, images and audio.
They are marked reference: we do not publish SophyAI prices or benchmarks until an end-to-end session has been completed. The official Meta repositories are gated and the weight license is tagged other, so its terms must be reviewed before commercial use.
Recommended configurations
For an enterprise deployment today, I would choose:
- 4×B200/B300 + vLLM for agentic coding with text and images;
- 8×H200 + SGLang when a 1M context is essential;
- 512 GB M3 Ultra + llama.cpp for private local use without cloud, accepting Q2 quantization;
- BF16 on 32×H200 only for research, fine-tuning or when full-checkpoint quality justifies the infrastructure.
The next step is to test the NVFP4 recipe on a Blackwell node, measure download, startup, memory, throughput and tool calling, and only then enable platform launch.