Back to Model
M

Muse Spark Open

Model Oct 16, 2026

Meta's 816B-total, 42B-active mixture-of-experts model for coding and agentic work, with multimodal inputs and community GGUF quantizations.

Visit primary link 3 sources attached
Hugging Face's oMLX Chat screenshot with Muse Spark Open selected and its reasoning effort control set to High.
Hugging Face's published oMLX interface screenshot showing Muse Spark Open and a reasoning control. It illustrates local software, not a generated answer or an AITrending throughput test. Resized without cropping. Source: Hugging Face / oMLX
multimodalself-hostedcoding-agents

What happened lately

Muse Spark Open timeline

Meta released Muse Spark Open on October 15, 2026. Its model card describes 816 billion total parameters, roughly 42 billion active per token, a one-million-token context window and text output. Supported input modalities depend on the serving runtime.

Hugging Face reports running an imatrix-calibrated two-bit quantization on a 512GB M3 Ultra at 18 tokens per second, with 304GB used by the model. The ggml-org card lists a Q2_K GGUF at 304GB. Active-parameter count is not the storage requirement: the routed experts still need memory. AITrending has not reproduced the performance measurement.

Reasoning effort ranges from low to xhigh and cannot be disabled in the documented local integration. Original weight access requires accepting Meta’s conditions. The custom Muse Model License includes usage and attribution requirements and a separate-agreement clause for qualifying Model-as-a-Service businesses above its revenue threshold; the clause explicitly exempts internal use and non-MaaS product features.