Meta Releases Open-Weight Multimodal MoE Model Muse Spark Open

Meta published Muse Spark Open on the Hugging Face blog, officially positioned as an open-weight frontier multimodal mixture-of-experts model. The source provides only this headline, with no specifics on parameter scale, active parameters, context length, modality coverage, benchmark scores, or license terms. For teams planning to self-host, the direct implication is one more downloadable frontier multimodal weight option. A MoE structure usually means only part of the parameters are activated at inference, so single-GPU VRAM pressure may be lower than a dense model of the same total parameter count. But open weights do not mean permissive licensing; whether commercial authorization is required or monthly active user limits apply depends on the model card and license file. Developers doing model selection on Hugging Face should first compare Muse Spark Open's model card, inference examples, and benchmark results against existing options before deciding whether to replace them or add it to a mixed routing setup.