Welcome Muse Spark Open: open-weight frontier multimodal mixture-of-experts model from Meta

Hugging Face’s Pedro Cuenca details Meta’s Muse Spark Open: roughly 816 billion total parameters, about 42 billion active per token, 256 experts with 8 selected per token, up to a 1 million token context, and inputs spanning text, images, audio, and video with text output. Checkpoints include BF16 and NVFP4 (Muse-Spark-Open-816B-A42B and the NVFP4 variant). Meta positions it as coding-optimized and agentic — plan, act, observe, self-correct — following the smaller Muse Glimmer dense model from August.

The post covers architecture (62 layers, perception and audio encoders), transformers distributed inference, llama.cpp quants (including 2-bit on Mac M3 Ultra at reported 18 tok/s), SGLang and vLLM NVFP4 recipes, MLX/oMLX quants, Inference Endpoints, and TRL LoRA fine-tuning across 64 H100s.

This is the deployment manual the launch needed. Meta’s claim of a U.S. open-weights frontier model only matters if people can actually run, quantize, and fine-tune it — and Hugging Face is doing that translation work in public.

The MoE math is the useful part for practitioners: all 816B stay in memory, but each token runs like ~42B. That means cluster-scale RAM with mid-size compute cost per token — NVFP4 on Blackwell and aggressive llama.cpp quants are how the rest of us get close. Demos tying the model to browser agents, Three.js asset loops, and voice-driven shopping show the multimodal bet isn’t decorative.

Our read: open multimodal agents just got a serious reference stack. The risk is the usual one — capable local agents plus tool use lower the bar for misuse even as they help builders. I’m watching which serving path (NIM-class cloud vs llama.cpp desk) wins real traffic.

Muse Glimmer is the smaller same-series dense model; both share tokenizer and vocabulary. Reasoning effort is adjustable but cannot be fully turned off on Muse Spark Open.