The short answer
● Want the most capable weights: Muse Spark Open — 816B total, 42B active.
● Want a permissive licence: Nemotron 3.5 Super VL under OpenMDW, not a bespoke model licence.
● Want to run it on hardware you own: both, but very differently — Mac Studios vs a single server GPU.
Meta and Nvidia both opened weights on 15–16 October. They are not competing for the same deployment, and the licence difference matters more than the benchmark difference.
Side by side
| | Muse Spark Open | Nemotron 3.5 Super VL |
|---|
| Total parameters | 816B | 120B |
| Active parameters | 42B | 12B |
| Architecture | Mixture of experts | Hybrid Mamba-Transformer MoE |
| Context window | 1M tokens | 1M tokens |
| Multimodal | Not stated at launch | Yes, vision-language |
| Licence | Muse Model License | OpenMDW |
| Target hardware | 2–4 paired M5 Ultra Mac Studios via MLX | BF16 and NVFP4 builds |
| Smaller variants | None announced | Bolt 74B-A9B, E4B edge |
The licence is the decision
Muse Model License is Meta's own document, not Apache 2.0 or MIT. Previous Meta model licences have carried use restrictions and scale thresholds, so anyone planning a commercial product on Muse Spark Open should read the terms against their specific use case before building. OpenMDW is a published open licence, which makes Nemotron the lower-friction choice for a product you intend to sell.
The hardware story is the other decision
Meta optimised Muse Spark Open for two to four paired M5 Ultra Mac Studios running MLX. That is a genuinely unusual target: a frontier-class open model on a desk rather than in a rack, with no per-token API cost and no data leaving the building. At 42B active parameters the arithmetic works, even though 816B total sounds impossible on that hardware.
Nemotron goes the other way. 12B active with vision and a million tokens of context is built for long agent loops where cost compounds across hundreds of steps, and the hybrid Mamba-Transformer backbone means the throughput advantage grows with context actually used rather than staying flat. Nvidia claims up to 5x comparable open models; that figure is unreproduced at publication.
How to choose
Building a commercial product → Nemotron, on licence clarity alone.
Running sensitive data on owned hardware → Muse Spark Open on Mac Studios, if the licence permits your use.
Running long multimodal agent loops → Nemotron, and test the throughput claim on your own context lengths.
Sources:
Meta Research ·
Hugging Face