Meta has launched Muse Spark Open, an open-weight multimodal artificial intelligence model featuring 816 billion total parameters, roughly 42 billion active parameters per token, and a context window reaching up to 1,048,576 tokens.
Text, images, audio, and video serve as inputs for the system, which generates text output and handles coding, tool usage, and other agentic workloads.
Muse Spark Open Specifications
Total parameters: Around 816 billion
Active parameters per token: Around 42 billion
Architecture: Mixture of Experts
Experts per MoE layer: 256
Experts selected per token: 8
Inputs: Text, images, audio, and video
Output: Text
Context window: Up to 1,048,576 tokens
Checkpoint formats: BF16 and NVFP4
816B Parameters With Mixture-of-Experts Design
A Mixture-of-Experts architecture powers Muse Spark Open, using 256 experts across each MoE layer while activating only eight for every individual token.
Through this setup, the system operates with approximately 42 billion active parameters per token while maintaining the overall capacity found in a system containing 816 billion parameters.
Meta makes the model available in BF16 and NVFP4 formats alongside quantized editions suited for different hardware configurations.
Supports a 1 Million-Token Context Window
Context lengths up to 1,048,576 tokens are supported by Muse Spark Open, divided across text, images, video, and audio inputs.
Visual and audio inputs use dedicated encoders within the model, and the audio component handles 24kHz audio at a rate near 12.5 tokens per second.
Tool calling and multiple reasoning tiers from low to high are also built into the system.
Meta Shows Coding and Browser Agents
Demonstrations by Meta showed Muse Spark Open managing various agentic assignments.
During one test, the model created and refined 3D objects using Three.js, relying on browser screenshots to assess its output and modify the code.
An additional demonstration featured the system operating a browser strictly from screenshots without accessing HTML or DOM information from the webpage.
Meta also exhibited a voice-activated shopping exercise where a customer detailed an item and the model applied its audio, visual, and agentic functions to locate a corresponding product.
As the biggest model within the Muse lineup from Meta, Muse Spark Open is built for large-scale multimodal, coding, and agentic tasks.
Can Run Locally on Apple Silicon
High-end Apple Silicon Macs can execute Muse Spark Open via llama.cpp, oMLX, and mlx-lm.
Testing a 2-bit edition on an M3 Ultra equipped with 512GB of unified memory consumed about 304GB while producing about 18 tokens per second.
A 3-bit edition used about 344GB and attained around 21 tokens per second during generation.
Even with substantial compression applied, the model demands more than 256GB of memory.
Nvidia GPUs Are Also Supported
Deployment of the BF16 edition spans multiple Nvidia H200 systems, whereas the NVFP4 edition accommodates newer Blackwell GPUs such as the B200 and GB300.
Frameworks such as SGLang, vLLM, and Hugging Face Inference Endpoints also integrate with the model.
