Meta's Muse Spark Open: what an 816B open-weight model means for on-prem
On October 15, Meta released the weights for Muse Spark Open: an 816-billion-parameter mixture-of-experts model with 42B parameters active per token, a context window of up to 1M tokens, and native image, video and audio input. Meta calls it the most capable US open-weight model to date. The weights are on Hugging Face and in the AWS, Google Cloud and Azure model catalogs.
Clients have already started asking whether they should run it. Short answer: probably not yet. Here is why, and when it might make sense.
The hardware bill
A mixture-of-experts model is cheap to run per token, because only 42B parameters fire each step. It is not cheap to hold. Every expert has to sit in memory, so you pay for all 816B:
| Precision | Weights alone (approx.) |
|---|---|
| 16-bit | ~1.6 TB |
| 8-bit | ~820 GB |
| 4-bit | ~410 GB |
That is before the KV cache. A 1M-token context is a headline number, but long contexts eat memory on top of the weights, so real deployments will use far less of it.
At 4-bit, the realistic options are a multi-GPU server with roughly 8× 80 GB cards or more, or a pair of high-memory unified-memory machines (people are already reporting it running across two 512 GB Mac Studios). Either works for a demo. Neither is the kind of box most small and mid-sized businesses want to buy, power and maintain for a team of 20.
Compare that with the one-GPU class of models: for most document Q&A, a well-built retrieval pipeline over a 30B-class model gets you closer to a good answer than a giant model with weak retrieval. We argued this in RAG that actually works, and nothing about this release changes it.
Read the license before you read the benchmarks
Muse Spark Open is open weights, not open source. It ships under Meta’s own Muse Model License. Reported terms include:
- Free commercial use up to $20M in revenue; above that you need a separate deal.
- Products must display “Built with Muse Spark Open.”
- Fine-tuning, distillation and redistribution are allowed, with conditions a standard open-source license does not have, including an 18+ age restriction.
For a regulated business, a revenue cap and a branding requirement are not small print. They are things legal has to sign off on. If you want Meta’s models with no strings, the separate Muse Glimmer (30B) is Apache 2.0, and it fits on far smaller hardware.
Meta’s own benchmarks also put Muse Spark Open behind the top Chinese open-weight models. For some clients, where a model comes from matters as much as how it scores, so this is a real option. It is not the best model on raw scores.
So who should run it?
- Probably yes: organizations already running multi-GPU inference, with data that cannot leave the building, and work that needs long multimodal context (video, audio, large document sets in one pass).
- Probably not: most teams whose real need is grounded Q&A over their own documents. A smaller model with good retrieval is cheaper, faster, and easier to keep running.
- Wait and see: everyone else. Quantized builds, inference-engine support and independent evaluations always land in the weeks after a release like this. Let them settle before you buy hardware.
The bigger story is that frontier-scale open weights keep arriving every month. The hard question for a business isn’t “can we get a strong model?” any more. It’s “what does it cost to run reliably, and what are we allowed to do with it?” That question is still best answered with real numbers for your own workload.