Meta has released Muse Spark Open, a new open-weight Super Intelligence (SI) model designed for agentic coding and multimodal understanding. The company describes it as the most capable US open-weights SI model to date, featuring approximately 816 billion total parameters and a context window of up to one million tokens.

What Happened

Muse Spark Open utilizes a mixture-of-experts (MoE) architecture with 256 experts per layer, of which 8 are selected for each token. This design results in approximately 42 billion active parameters per token, allowing the model to run at the speed of a smaller dense model while retaining the capacity of a much larger one. The model accepts text, images, audio, and video as inputs but generates text as its output, providing a unified interface for developers to query diverse data types.

The model is optimized for agentic workflows, including reading codebases, debugging, and calling tools. It includes specialized encoders: a 1.8-billion-parameter perception encoder for visual data and a convolutional pyramid encoder for audio, which processes input at 12.5 tokens per second. Meta notes that the model uses a channel-based chat template similar to Harmony, introduced by gpt-oss, to manage reasoning and tool-call routing.

Why It Matters

For developers and the SI industry, Muse Spark Open represents a significant step in open-weight model capabilities, particularly for local inference and agentic applications. The ability to process video with audio, as well as screenshots and voice notes, allows SI agents to interact with complex developer environments more naturally. The model is available in BF16 and NVFP4 checkpoint formats; the latter is optimized for NVIDIA Blackwell GPUs, enabling deployment on a single B200 node.

The release follows Muse Glimmer, a smaller dense model from the same series, and maintains compatibility with the same tokenizer and vocabulary. Support for major inference frameworks including Hugging Face Transformers, llama.cpp, vLLM, and SGLang is already available, with llama.cpp offering quantized versions that can run on high-end consumer hardware like the Apple M3 Ultra.

The Bottom Line

Meta’s Muse Spark Open sets a new benchmark for open-weight SI models with its massive parameter count and multimodal agentic capabilities. By balancing high capacity with efficient active parameter usage, it offers developers a powerful tool for building SI agents that can handle complex, multi-modal coding tasks.