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

Hugging Face's launch post for Meta's Muse Spark Open is mostly a practical guide to running and adapting the model (Hugging Face). It shows working setups for transformers, llama.cpp, vLLM, SGLang, Apple's MLX and Hugging Face's own hosted endpoints, plus a fine-tuning recipe. The architecture explains why setup matters so much: only about 42B of the model's 816B parameters work on each token, but all of them must be held in memory, which is why the full-precision weights need multiple GPU servers.

The smallest setup the authors report is a single Mac with an M3 Ultra and 512GB of memory, and only with aggressive compression. A 2-bit llama.cpp version used about 304GB and generated about 18 tokens per second; a 3-bit MLX version took about 344GB and generated about 21. The post warns that even 3-bit needs more than 256GB of RAM, and that the long context has its own memory bill: a full 1M-token conversation adds roughly 266GB of cached attention data at BF16.

Two details change everyday use. The model always reasons before it answers: reasoning cannot be turned off, the default effort is the highest setting, xhigh, and the post suggests dropping to low to spend fewer tokens on easy tasks. And for adaptation, the authors trained small LoRA adapters on the attention layers while leaving the experts frozen, a run of 500 steps that took about two hours on 64 H100 GPUs.