Meta Muse Glimmer runs offline on one GPU, but needs 24 GB VRAM

Meta released Muse Glimmer on August 10, a language model with roughly 30 billion parameters. What stands out is not the model itself so much as the license and the target machine. The weights sit on Hugging Face under Apache 2.0, so you can download them, modify them and use them commercially. And this is not built to run in a data centre, but on a single graphics card under your desk.

The model comes from Meta Superintelligence Lab and is distilled from the larger Muse Spark, meaning a big model taught a small one how to answer. It takes text and images as input but only produces text, handles more than 100 languages and works with a context window of over 131,000 tokens. Its knowledge cuts off on January 4, 2026. Meta's most recent releases were the Muse Code agent and the Muse Image generator.

24 GB of VRAM is the price of entry

At full precision the model would need 64 GB of video memory. Meta compresses it to roughly 4-bit through quantization, a coarser way of storing numbers that takes up far less room. That shrinks the language part to under 20 GB, which leaves space for the vision encoder and the cache for the running conversation.

Two ready-made variants ship with it. K-Quant-Dynamic targets 32 GB and loses 0.2 percent accuracy on average across 15 benchmarks, while K-Quant-17GB fits into 24 GB and loses 1.0 percent. In practice that means a GeForce RTX 5090, RTX 4090, RTX 3090 or a Mac with an M4 Max. A mid-range card with 12 GB is out. If you have less memory, our guide covers smaller models for your own laptop, and even the iPhone now runs a usable language model.

One accelerator makes it three times faster

To keep a local model from feeling sluggish, Meta ships a helper called DFlash. It proposes 16 tokens at once, and the large model checks the whole block in a single pass and only corrects what is wrong. By Meta's own measurement, throughput on an RTX 5090 climbs from 74.9 to 233.4 tokens per second, a factor of 3.1. A MacBook with an M5 Max goes from 26.6 to 50.2 tokens, the M4 Max from 23.7 to 37.8.

AMD published its own numbers the same day. A mini PC with a Ryzen AI Max+ 395 manages up to 24 tokens per second and a Radeon AI PRO R9700 up to 53, measured on Windows with llama.cpp. AMD labels the figures as preliminary.

Strong at planning, weaker at the interface

Meta compares Muse Glimmer against Google's Gemma4-31B and Alibaba's Qwen3.6-27B. When the model calls tools and plans across several steps, it leads clearly. It scores 75.5 on MCP Atlas against 54.2 and 62.5, and 74.6 on DeepSearch QA against 61.7 and 71.1.

Qwen wins elsewhere. On driving an actual desktop interface the score is 65.9 to 75.6, in the terminal 51.7 to 60.7. And on the Siren AgentDojo security test, which measures how easily a model can be manipulated by instructions hidden inside outside documents, Muse Glimmer's 28.4 percent attack success rate beats Qwen but trails Gemma4 at 25.6 percent. Worth keeping in mind before you point the model at your own files and mail.

How to get it

Four packages are waiting on Hugging Face: the base model with full BF16 weights, GGUF files for llama.cpp, ExecuTorch builds for Apple devices and the DFlash helper. The easiest route is LM Studio, where the model turns up in search. AMD recommends more than 32 GB of video memory or an equivalent allocation of system memory for that. If you have less, providers such as Unsloth offer smaller quantizations, but you then have to offload part of the model into ordinary system memory, which costs noticeable speed. The community moved fast. Less than a day after release there were already 57 additional quantizations and six fine-tunes sitting alongside it. Trying it out in a browser the way you could with Kimi K3 is not an option here. Without the right hardware, Muse Glimmer stays out of reach.