avoid.worksNotes, news & experiments in AI.
← Back to news
NEWS BRIEF / OPEN WEIGHTSOct 15–16, 2026

Two giant open models, and a practical catch

Muse Spark Open and Beam can be downloaded. Running either one yourself is another matter.

avoid.works · Source summary and editorial context · Published Oct 18, 2026

Hover, click, or tap a dotted-underlined term for a quick explanation and a link to read more.

Think of getting the plans for a very large machine. You may be allowed to build it, but you still need the room and equipment. That is the lesson in two October releases: Meta's Muse Spark Open, released October 15, and Reflection's Beam, released October 16. Their model weights, the learned numbers stored in files, are available to download. Neither is a practical fit for an ordinary laptop.

What arrived?

ModelWhat it can take inScale
Muse Spark OpenText, images, audio, and video; writes textAbout 816 billion total parameters, with about 42 billion active for each token
BeamText; writes textAbout 501 billion total parameters, with about 23 billion active for each token

Both use a mixture of experts: the model selects some groups of its learned numbers for each piece of text it processes. That can reduce calculation per answer. It does not remove the need to store all those groups somewhere while the model runs.

All model weights must be stored in memory, while only selected expert groups are used for each token; fewer active groups do not make the whole model small.
The dark blocks stand for selected experts used on one token. The lighter blocks still need memory.

Can I run one on my computer?

Hugging Face reports that a tested Muse Spark Open MLX build used roughly 344 GB of memory on an M3 Ultra with 512 GB of unified memory. Beam's smaller NVFP4 checkpoint is about 280 GB; Hugging Face recommends four B200 GPUs for comfortable serving. Those are specific published setups, not speed or cost promises. They also put both models far beyond a 64 GB MacBook Pro. Hosted providers are the simpler way to try them without owning that hardware.

Does “open” mean the same thing for both?

No. Muse Spark Open's weights are available under Meta's Muse Model License, which includes a usage policy. Beam's release says it uses the Apache 2.0 license. Always read the license before building or redistributing a product. “Open weights” describes access to model files, not a promise that training data, every development detail, or free computing are included.

For a home machine, start with the model finder or our guide to memory and speed. Compare smaller models on your own tasks instead of assuming the largest available model is the best choice.

Sources & further reading