Muse Spark Open: Meta's 816B Open Model Runs Like a 42B One

Meta’s new open-weight model has about 816 billion parameters, yet it computes each token with only about 42 billion of them, and that gap decides who can actually run it. Hugging Face’s launch write-up of Muse Spark Open describes a model Meta trained for agentic coding: reading a codebase, writing and debugging code, and calling tools, with images, audio and video accepted as input. For teams that want a self-hosted coding agent instead of a paid API, the question is less whether it is good and more what hardware it demands.

What Happened

Muse Spark Open is the full-size model in Meta’s Muse series, following August’s Muse Glimmer, a small dense model built for local use. According to The Decoder, Meta calls it the most capable US open-weight model and says it scores well ahead of Qwen 3.8, GLM 5.3 and Kimi K3 in several categories on benchmarks including JobBench and MCP-Atlas, while falling well short of Meta’s paid Muse Wonder Full models. Those are Meta’s own figures.

The weights are on Hugging Face in BF16 and NVFP4 formats, with a context window of up to 1 million tokens. The license is free to use but requires a “Built with Muse Spark Open” label, and providers earning more than $20 million a year who want to offer it as an API service need a separate agreement with Meta.

Why It Matters

The architecture explains both the appeal and the cost. Each mixture-of-experts layer holds 256 experts, and a router picks 8 per token, so the model runs at roughly the speed of a 42B model. All 256 experts must stay in memory, though, so deployment needs the memory of the full 816B model. The BF16 checkpoint is about 1.5 TB. In transformers, it runs across 4 nodes of 8 H200 GPUs, while the 4-bit NVFP4 version fits on a single B200 node.

Long context has its own bill. Hugging Face calculates the KV cache at about 250 KB per token in BF16, or about 266 GB for a full 1-million-token context, on top of the weights.

Local use is possible at the high end. Hugging Face tested a 2-bit quantization on a Mac M3 Ultra with 512 GB of memory at 18 tokens per second, using 304 GB for the model, and a 3-bit MLX quantization at about 21 tokens per second. Reasoning can’t be switched off; the default effort is xhigh, and Hugging Face suggests setting it to low for easier tasks to save tokens.

What This Means for Builders

The launch post is practical about agent use. It shows tool calling through a channel-based chat template, and demos pairing the model with a coding agent that writes Three.js code, renders it, screenshots the result and feeds the images back to the same model for critique. That loop, where the model reads its own visual output, is a concrete use for the multimodal input. Fine-tuning takes serious hardware: Hugging Face’s LoRA run used 64 H100 GPUs for about two hours. Stack Archive’s earlier look at Meta’s Muse Glimmer and local agentic AI covers the small end of the same family, which shares Muse Spark Open’s tokenizer.

Source: Hugging Face

Content created with AI assistance and reviewed for accuracy.