What shipped
| Reflection Beam | Meta Muse Spark Open |
|---|
| Released | 16 October 2026 | 15 October 2026 |
| Parameters | 501B | 816B total |
| Licence | Apache 2.0 | open weights |
| From | Reflection AI, a startup | Meta |
| Tooling | Hugging Face Transformers v5.20.1 | Hugging Face Transformers v5.20.0 |
A day apart, and between them the two largest open-weight releases of the month.
The licence is the real difference
Beam is Apache 2.0. That is an OSI-approved licence with no field-of-use restriction, no user-count threshold and no naming requirement. You can build a commercial product on it without reading the terms twice.
Muse Spark Open is described as open weights, which is a weaker and less precise claim. Meta's open releases have historically carried conditions - and whatever the specific terms, "open weights" and "Apache 2.0" are different commitments. If you are building something you intend to sell, read Meta's licence text before you build; you do not need to read Beam's.
Scale against what you can serve
816B against 501B looks decisive and mostly is not, because neither fits anywhere convenient. Both are models you serve on a cluster or rent from a provider, and at that point the question is not which has more parameters but what each costs per token in practice.
DeepSeek V4.1-Flash is a useful reference here: 552B as a mixture-of-experts, served at $0.30/$1.20 per million. Parameter count alone tells you little about serving cost - architecture and active parameters per token decide it, and neither release has published figures that make a clean comparison possible.
Where they come from matters
Reflection AI is a startup that Nvidia is reported to be in talks to acquire, invest in or acquihire. Nvidia already led a $2 billion round last October and has roughly $800 million in; the lab was valued near $25 billion pre-money in April and holds a SpaceX compute deal worth up to $6.3 billion. If that acquisition closes, the open-weights policy of a Reflection inside Nvidia is an open question - though Nvidia selling chips has a clearer interest in open weights than most buyers would.
Meta detailed Prometheus this week: a gigawatt-scale site in New Albany, Ohio, eventually targeting about a million GPUs, explicitly for training the Muse family. Muse Spark Open is a release from a pipeline that is very obviously going to keep producing, which is worth something if you are adopting a model family rather than a model.
Pick one
- Commercial product, licence certainty - Beam. Apache 2.0 removes the question entirely.
- Largest available open model - Muse Spark Open, at 816B.
- Coding - Beam. Reflection has consistently positioned it there, and it already appears in open-weight coding comparisons.
- Betting on a family rather than a model - Muse Spark Open. Meta has a gigawatt of training capacity pointed at it.
- Serving it yourself on limited hardware - neither. Look at DeepSeek V4.1-Flash, which is smaller, MIT-licensed and has a published serving stack now that DeepInfer is open source.
Verdict
Beam is the better default because Apache 2.0 is the strongest licence commitment of the two, and licence risk is the one thing you cannot engineer around later. Muse Spark Open is the bigger model from the better-resourced lab, and if your work benefits from raw scale that is a real argument.
The honest caveat: neither has published the benchmark results needed for a capability verdict, and both are large enough that serving economics will decide the question for most teams before quality does.