Meta's new Muse Glimmer AI model is supposed to run on a laptop with 18GB of memory. We tried it on a normal office laptop with no dedicated graphics card, and it took 12 minutes to respond to a two-sentence coding query. That was the only attempt that worked. The two before it failed outright.
Muse Glimmer is a free, open-weight AI model released by Meta on August 10, 2026. Unlike ChatGPT and Claude, there is no need to use it through a website. Instead, you download the model to your computer, where it runs offline, and nothing you type leaves your device.
The logic behind it was laid out in a 14-page essay released by Mark Zuckerberg the same day, titled "The Future Is for Everyone." In it, he argues for distributing AI across millions of personal devices rather than keeping it inside a few companies' data centres.
Meta pitches the model against Alibaba's Qwen3.6 27B and Google's Gemma 4 31B, the two rival free models built for the same job, and its own benchmark table shows Qwen winning several of those matchups.
Damilare Odedina
Muse Glimmer speed test: 12 minutes for one answer
We requested Muse Glimmer to generate a small program written in Python that would determine if a number is prime, and explain in two sentences how efficient that code is. Simple enough assignment for a first-year CS student.
The first attempt took 154 seconds just to produce its opening word, then crashed with a "network error" message. The second spent 565 seconds on calculations and failed after roughly ten minutes without generating any response at all.
The third attempt was successful, though. Muse Glimmer provided the required answer at the rate of 0.9 words per second, taking 711 seconds in total, almost 12 minutes. Compare it to the few seconds needed by ChatGPT or Claude for their answer.
The answer was decent: correct program code, precise explanation, appropriate length of the text. Quality was not the problem here. The long waiting time was.
Why Muse Glimmer is so slow without a graphics card
The limiting factor is memory throughput, not computing power.
Subscribe for free to continue reading this article
Subscribe SubscribeAlready have an account? Log in
