Meta AI Releases Muse Glimmer: A 30B Open-Weights Agentic Model That Runs on One Consumer GPU

Meta has put out Muse Glimmer, a 30-billion-parameter multimodal model distilled from Muse Spark and released under Apache 2.0. The pitch is blunt: keep it running all the time, keep it local, and keep it cheap enough that a single consumer GPU can handle it.

That matters because a model this size normally wants more than 55 GB of memory at full precision. Meta squeezes it down to roughly 4-bit, then pairs that with block-level speculative decoding so the model can answer quickly enough to fit inside an agent loop. The result is meant to run without a network call, on one GPU or even a Mac.

The distribution looks built for self-hosting from day one. Meta says the Hugging Face collection includes BF16 weights, GGUF k-quants, ExecuTorch builds, and the DFlash drafter. It also points to use cases that care about being offline or air-gapped: healthcare, legal, financial services, defense and public sector, manufacturing, and field service. The list of applications leans heavily toward agent work — screenshot-reading desktop agents, coding agents, schema-based function calling, document and chart understanding, synthetic data, and LLM-as-a-judge evaluation.

The model itself is a dense causal transformer with a dedicated perception encoder. It handles text and images, not audio, and video is treated frame by frame. Meta says the context length goes beyond 131,072 tokens, the vocabulary is 202,048 tokens, and the knowledge cutoff is January 4, 2026.

On performance, Meta is making two promises at once: fit and speed. The 4-bit compression brings the language model under 20 GB, leaving room in a 24 GB or 32 GB setup. It ships two quantized builds, with K-Quant-17GB aimed at 24 GB VRAM and K-Quant-Dynamic aimed at 32 GB. DFlash predicts 16 tokens in one forward pass, and Meta says that raises throughput on an RTX 5090 from 74.9 to 233.4 tok/s. Apple’s M5 Max goes from 26.6 to 50.2 tok/s, and M4 Max from 23.7 to 37.8 tok/s.