Meta has released Muse Spark Open, an open-weight multimodal AI model with 816 billion total parameters, around 42 billion active parameters per token, and a context window of up to 1,048,576 tokens.
The model supports text, images, audio, and video as input and generates text as output. It is designed for coding, tool use, and other agentic tasks.
Muse Spark Open Specifications
| Specification | Muse Spark Open |
|---|---|
| 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
Muse Spark Open uses a Mixture-of-Experts architecture with 256 experts in each MoE layer, but only eight are activated for each token.
This allows the model to use roughly 42 billion active parameters per token while retaining the capacity of the much larger 816-billion-parameter system.
Meta is offering the model in BF16 and NVFP4 formats, along with quantized versions for different hardware.
Supports a 1 Million-Token Context Window
Muse Spark Open supports up to 1,048,576 tokens of context, shared between text, images, audio, and video.
The model includes separate encoders for visual and audio input. Its audio system processes 24kHz audio at around 12.5 tokens per second.
It also supports tool calling and several reasoning levels ranging from low to high.
Meta Shows Coding and Browser Agents
Meta demonstrated Muse Spark Open handling several agentic tasks.
In one example, the model generated and improved 3D objects in Three.js, using browser screenshots to evaluate its work and edit the code.
Another demonstration showed the model controlling a browser from screenshots without receiving the webpage’s HTML or DOM data.
Meta also demonstrated a voice-driven shopping task in which a user described a product and the model used its audio, visual and agentic capabilities to search for a matching option.
Muse Spark Open is the largest model in Meta’s Muse family and is designed for large-scale multimodal, coding and agentic workloads.
Can Run Locally on Apple Silicon
Muse Spark Open can run through llama.cpp, oMLX, and mlx-lm on high-end Apple Silicon Macs.
A 2-bit version tested on an M3 Ultra with 512GB of unified memory used around 304GB and generated roughly 18 tokens per second.
A 3-bit version used around 344GB and reached about 21 tokens per second during generation.
Even with heavy compression, the model requires more than 256GB of memory.
Nvidia GPUs Are Also Supported
The BF16 version can run across multiple Nvidia H200 systems, while the NVFP4 version supports newer Blackwell GPUs such as the B200 and GB300.
The model also works with frameworks including SGLang, vLLM and Hugging Face Inference Endpoints.
The post Meta Muse Open is a Top New Open Weight AI Model appeared first on ProPakistani.