VentureBeat reported that Muse Spark Open weights are available on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. API routes include Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI, with context limits varying by host. Together AI lists $1.25 per million input tokens and $4.25 per million output tokens.
Self-hosting is demanding: the model card requires at least 700GB of accelerator memory for NVFP4 and 2TB for BF16. Meta's demonstration used two M5 Ultra Mac Studios with up to 512GB each. Serving software supported includes SGLang, vLLM, TensorRT-LLM, Transformers, MLX, llama.cpp and Unsloth.
Source: VentureBeat