Muse Glimmer Targets Consumer VRAM With a 30B Multimodal Stack

What happened

Meta Superintelligence Labs released Muse Glimmer, a dense 30 billion parameter vision-language model, with weights published as meta-models/Muse-Glimmer-30B on Hugging Face Hub. The package ships with day-zero integration for transformers, llama.cpp, vLLM, and Inference Endpoints under an Apache-2.0 license.

Context

Open-weight vision-language models have steadily moved toward agentic reasoning, but deployment historically required multi-GPU clusters or cloud endpoints. Muse Glimmer targets that gap by compressing the decoder to under 20 GB in 4-bit quantization and routing attention for consumer accelerators.

How it works

The architecture splits a 2B ViT-style perception encoder from a 28B text decoder. The encoder runs 50 transformer layers with GELU MLPs on video capped at 96 evenly sampled frames at 2 fps. The decoder applies hybrid attention across 52 total layers: three sliding window layers followed by one full attention layer, repeated 13 times. It uses grouped-query attention with a 16-to-1 head ratio and Q-K normalization. An optional DFlash speculative decoding drafter runs alongside generation. Inference targets NVIDIA, AMD, and Intel via automatic device mapping.

Our read

The 24 GB to 32 GB VRAM envelope is the headline number, but it conceals a tight boundary during agentic loops. The encoder, KV cache, and DFlash drafter share that memory ceiling with the quantized decoder, which means context switching or tool-calling will spike utilization quickly. The 96-frame limit at 2 fps also forces chunking for any long-form scene analysis; this model is built for shot identification and metadata tagging, not narrative comprehension. Day-zero support lists standard inference runtimes but omits ComfyUI entirely. Since local video production pipelines run on ComfyUI daily, bridging Muse Glimmer will require a custom node rather than dropping it into an existing workflow. DFlash speculative decoding does change the latency profile for prompt iteration, which matters when drafting scripts or refining vision queries offline. The model trades clip length and immediate pipeline integration for offline multimodal capability without cloud dependency.

What this changes

Run the quantized stack on a single consumer GPU for local ingestion. Point vLLM to process shot breakdowns without API costs. Build a ComfyUI node to route video placeholders, since day-zero runtimes omit it. Accept the 96-frame cap and split longer footage. Enable DFlash for faster prompt iteration.

License

Released under the Apache-2.0 license, which permits commercial use, modification, distribution, and private fine-tuning with attribution. The terms do not restrict redistributing generated metadata or building commercial pipelines around the weights.

Key takeaways

• Muse Glimmer packs a 30B multimodal model into a 24–32 GB VRAM envelope by splitting a 2B vision encoder and quantizing the 28B decoder.
• Hybrid attention with grouped-query scaling and an optional DFlash drafter targets consumer accelerators without cloud dependency.
• Day-zero support covers transformers, llama.cpp, vLLM, and Inference Endpoints; ComfyUI integration requires a custom node.
• The 96-frame limit at 2 fps restricts long-form analysis to segmented workflows.

Sources

1. Meta is back with Muse Glimmer: local, agentic, multimodal, and open source — tier 1
2. Meta Muse Glimmer – open weights 30B local coding model — tier 3
3. Introducing Muse Glimmer: an open-weight model optimized for always-on local agent workflows — tier 3