Meta released the weights for Muse Spark Open on October 15 and is calling it the most capable US open-weight model to date. The specs: natively multimodal, 816B total parameters with 42B active (mixture-of-experts), a 1M-token context window, and built for coding and agentic work.
The local angle
This is the part I care about. Meta says it is the first model in its class to run on a desktop — specifically two Apple M5 Ultra Mac Studios using the latest MLX. Each M5 Ultra can give its GPU up to 512 GB of unified memory, so two machines hold the whole model, and Meta says the 1.2 TB/s of memory bandwidth is enough to read the small slice each token needs.
Let’s be honest about “local” though. Per the Hugging Face model page, the weights come in BF16 (about 1.65 TB) and NVFP4 (about 537 GB). That is not a gaming-laptop model. It is “a serious home lab” model. Still, frontier-class open weights you can run off the cloud is a big deal for privacy and for anyone building agents they actually control.
Where it runs
Runtimes listed: SGLang, vLLM, TensorRT-LLM, Transformers, Apple MLX, llama.cpp and Unsloth, with optimizations for AMD, Apple, Intel and NVIDIA. Hosted APIs include Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Meta says it tested on-device agent workflows with Muse Code, including shipping real iOS changes to the open-source Ice Cubes Mastodon client.
One catch: the license is the Muse Model License with limited commercial terms. Free for many developers — read it before you build a business on it.
Open weights keep getting bigger and better. The hardware just has to keep up.
Have a blessed day! And Happy Hacking!