Meta releases Muse Spark Open, 816B-parameter coding model for Mac desktop

Meta has released Muse Spark Open, an open-weight multimodal model designed for coding and agentic tasks. The model features 816 billion total parameters with 42 billion active per token, a 1 million token context window, and a mixture-of-experts architecture with 256 experts per layer (8 selected per token). It processes text, images, audio, and video inputs and outputs text, optimized for complex software engineering workflows including multi-file refactors, debugging, web browsing, and tool integration.

A key differentiator is Muse Spark Open's capability to run locally on consumer hardware. Meta optimized the model to run on Apple's Mac Studio equipped with the M5 Ultra chip, leveraging Apple's unified memory architecture and MLX framework. Two M5 Ultra Mac Studios with up to 512 GB unified memory per machine can hold the entire model, providing approximately 1 terabyte of combined memory. The model runs on major inference engines including SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth with optimizations for AMD, Apple, Intel, and NVIDIA hardware.

Meta demonstrated the model's practical capabilities by deploying it to real workflows: running iOS fixes and new features for Ice Cubes (an open-source Mastodon client) and automating video editing tasks in DaVinci Resolve, including frame-by-frame review and vertical cropping. The model went through safety evaluation aligned with Meta's Superintelligence Scaling Framework before release and includes additional safety post-training for open-weight deployment.

Muse Spark Open is available immediately through multiple channels: model weights can be downloaded from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. Hosted inference is available via Together AI, Fireworks AI, Groq, Baseten, Modal, Nebius, and OpenRouter. The model is released under the Muse Model License with limited commercial terms allowing free use by many developers. The BF16 checkpoint is approximately 1.5 terabytes in safetensors format; a 4-bit NVFP4 variant is optimized for NVIDIA Blackwell GPUs. This follows Meta's August 2026 release of Muse Glimmer, a smaller dense model from the same series.