Meta released the weights for Muse Spark Open on October 15, 2026. The company calls it "the most capable US open-weight model to date." It is built for coding and agentic work, and Meta says it is the first model in its class that can run on a desktop setup of two Apple M5 Ultra Mac Studios.
The key specs
- Parameters: 816 billion total, 42 billion active per token
- Architecture: Mixture-of-Experts Transformer with a Perception Encoder; 256 experts per MoE layer, 8 selected per token
- Context window: up to 1 million tokens
- Inputs: text, images, video and audio.
- Output: text
- Languages: trained on data from more than 100 languages
- Formats: BF16 and NVFP4 checkpoints
Meta says the model excels at long-horizon tasks, complex software engineering and multi-step reasoning. The model card recommends a high or xhigh reasoning effort for complex problem solving, coding and agentic tasks.
What you need to run it
According to the model card on Hugging Face, running the weights on your own GPUs requires at least 700GB of VRAM for the NVFP4 version and 2TB for BF16. Supported inference frameworks include SGLang, vLLM, Hugging Face, TensorRT-LLM, Unsloth and Llama.cpp. On Apple silicon, the model is optimized to run on two Mac Studios with M5 Ultra using the MLX framework.
Meta also ships custom Helion GPU kernels in SGLang, which it says deliver 1.03x to 3.95x kernel-level speedups on NVIDIA and AMD GPUs.
Where to get it
The model is available through API providers Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Weights can be downloaded from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models.
The license
Muse Spark Open is released under Meta's custom Muse Model License, not a standard open-source license. The model card notes that this open-weight version received additional safety post-training, and Meta recommends deploying it as part of a broader system with extra guardrails suited to the use case.
Part of the Muse family
Muse Spark Open is the full-size model in Meta's Muse series. In August, Meta released Muse Glimmer, a smaller dense open-weight model from the same series. Both share the same tokenizer and vocabulary, according to Hugging Face.
