Meta · Muse

Muse Glimmer

Muse Glimmer is Meta’s 30-billion-parameter open-weight multimodal model, distilled from Muse Spark and aimed at local agent workflows. Weights are under Apache License 2.0.

Status: Current, open weights

Training

Dense, decoder-only multimodal transformer with a built-in vision encoder. Trained from Muse Spark’s outputs rather than from scratch (Meta docs). Full dataset mix and FLOPs are not published on the docs cited here.

Inference

Text and image in, text out. Default context is 128K tokens; longer contexts are supported per docs. Runs locally via vLLM, SGLang, llama.cpp, and ExecuTorch; partner hosts listed on docs include Ollama, LM Studio, and Together. No API key is required for self-host.

Context

Default context is 128K tokens. Exact maximum is on the live Glimmer docs.

Methods

Step-by-step reasoning before answer (docs). Distilled from Spark. Prompting and chat template: Meta Glimmer prompting guide. This is weights you run, not a chat-API SKU on the Meta Model API in the same way as Spark.

Benchmarking

A public third-party score table for Glimmer is not on the pages cited here. Prefer Meta research blog or docs if they publish dated evals later.

Hardware

Quantized GGUF builds are aimed at a single ~24 to 32 GB GPU envelope (Meta Glimmer blog / quantization docs). Exact training cluster size is not published.

Access

Hugging Face: meta-models/Muse-Glimmer-30B (bf16), meta-models/Muse-Glimmer-30B-GGUF, plus ExecuTorch / draft checkpoints per get-the-model docs. Apache 2.0.