Meta has put a very large multimodal model into the open. Muse Spark Open, described in a Hugging Face write-up on 15 October 2026, accepts text, images, audio and video in a single context of up to 1 million tokens and replies in text. Meta pitches it as the most capable US open-weight model so far and trained it for agentic coding: reading codebases, debugging, calling tools and working through multi-file refactors. For teams that want frontier-scale weights they can inspect and host themselves, it widens the choice considerably.
The architecture is a mixture of experts. Of about 816 billion parameters, roughly 42 billion are active for each token; each expert block holds 256 experts, of which 8 run per token, and the experts account for about 97% of the total. A 1.8 billion parameter vision encoder handles images and video, sampled at 2 frames per second by default, and an audio encoder turns every 80 milliseconds of sound into one token, so 10 seconds of audio costs 125 tokens. Reasoning effort can be set from low to xhigh but cannot be switched off, and xhigh is the default, so simple jobs may burn more tokens than they need.
The hardware bill is the practical barrier. The BF16 checkpoint is about 1.5 TB and the Transformers example spreads it across four nodes of eight H200 GPUs. A 4-bit NVFP4 version fits on a single B200 node, and a full 1 million token context needs about 266 GB of key-value cache on top of the weights. Community quantisations bring it within reach of a 512 GB Mac Studio, with a 2-bit build using about 304 GB and generating about 18 tokens a second. Most people will reach it through hosts; our review of Muse Spark Open lists OpenRouter at $1.25 per million input tokens and $4.25 per million output.
It arrived in the same week as Reflection's Beam, a 501B coding model under Apache 2.0, and together they suggest US labs are competing again at the top of open weights. Native audio and video input is what sets Muse Spark Open apart; most open rivals handle text and images at most. It also sits beside Meta Muse, the company's consumer assistant, which we review separately.
Some questions remain. The Hugging Face post includes no benchmark scores and no comparison with other models, so the "most capable" claim rests on Meta's word for now. The licence is not Apache or MIT: our review notes that the Muse Model License requires attribution and sets a $20M revenue threshold for offering it as a service. vLLM currently supports only text and image input. The post also contains small errors in its code examples, a sign that tooling is still settling. Independent evaluations on coding and multimodal tasks should come before any serious commitment.