Source: Muse Spark Is Now an Open-Weights Model (PyTorch blog)

Meta's announcement of Muse Spark Open: an 816B/42B-active multimodal MoE with 1M context under the custom Muse Model License.

Contents

Summary

Announces open weights for Muse Spark, Meta’s post-Llama model family, with extra safety post-training for standalone use, Helion kernels for SGLang and TorchTitan fine-tuning support.

Key claims

  • 816B total / 42B active parameters, 1M-token context, natively multimodal (image, audio, audio-video, documents).
  • Released on Hugging Face under the Muse Model License (not an OSI licence).
  • Shares a base model with the hosted Muse Spark 1.3; coding- and agent-optimised.
  • Additional safety post-training targets robustness when deployed without system guardrails.
  • Optimised on SGLang and vLLM across NVIDIA, AMD and Intel; TorchTitan supports SFT, LoRA and RL.
  • Meta previously released Muse Glimmer, a small dense open model.

Feeds

Open-weight models

Suggest changes →