← Model indexMODEL PROFILE / muse-spark-open

Meta / current

Muse Spark Open

Meta's open-weight 816B mixture-of-experts model, taking over as its strongest downloadable checkpoint.

Context1M
ModalityText · Image · Video · Audio
AccessWeights · Partners
ReasoningControllable

Where it fits.

A sparse MoE with 42B active parameters, one-million-token context and multimodal input, aimed at coding and agentic self-hosting. Released October 15, 2026 as meta-models/Muse-Spark-Open-816B-A42B with 816B total and 42B active parameters, a 1M-token context and multimodal input. Weights ship in BF16 (~1.65TB) and NVFP4 (~537GB) behind manual Hugging Face approval, with day-one support from SGLang, vLLM, TensorRT-LLM and llama.cpp.

02

Core capabilities

  • 816B total and 42B active sparse MoE
  • 1M-token context with multimodal input
  • Text-only generation for coding and agents
  • Runs on NVIDIA Blackwell, AMD Instinct, Xeon 6 and Mac Studio
03

Best-fit workflows

  • Self-hosted coding agents
  • Sovereign or air-gapped deployments
  • Provider-hosted open inference
04 / OPERATIONS

From evaluation
to production.

Deployment notes

Self-host only where control or residency justifies the memory footprint. Start from NVFP4 checkpoints, benchmark target context lengths rather than empty-model fit, and review the Muse Model License before any hosted resale.

Watch before choosing

Uses the Muse Model License rather than Apache 2.0, requiring attribution and imposing a revenue clause on hosted resellers. Hardware demand is severe: BF16 operation approaches 2TB of VRAM, and the license is not OSI-approved, carrying attribution duties plus a model-as-a-service clause above $20M group revenue.

05

Family & versions

Specifications and positioning checked against provider documentation.

Open official source↗