Meta’s Muse Glimmer fits on a laptop
The New Stack Amanda Caswell ● Covered by 3 sources
Meta’s Muse Glimmer is a 30B model built to run on a laptop. It turns cloud-trained brains into a local agent, with less data leaving the machine.
Based on reporting by The New Stack, Amanda Caswell — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Meta shipped Muse Glimmer on Monday, and the hook is simple: this is a 30-billion-parameter open-weight model built to run agentic workflows on local hardware. It’s on Hugging Face now. The bigger point is that Meta has turned its Muse Spark model into something small enough to act like a local helper instead of a remote cloud brain.
The path from Spark to Glimmer is the interesting part. Meta trained Glimmer on Spark’s outputs with logit distillation, then followed that with longer-context training that pushed harder on agents and richer reasoning traces. After that came the usual mix of supervised fine-tuning, reinforcement learning, and on-policy distillation across coding, reasoning, and agentic tasks. And then Meta added a lightweight secondary model so the system could speed through longer jobs.
This is more than a model release. It is a deployment pipeline, built around the idea that a strong central model can teach a smaller one to work nearby, on user hardware, while the larger model still handles the heavy lifting. That reduces cloud dependence for tasks like local coding work, where a company may not want source code leaving the machine. But it also adds more moving parts for developers to track, test, and rebuild when the parent model changes.
Meta says Glimmer can run on a Mac or PC with a single consumer GPU, though full precision needs more than 55GB of memory. The 4-bit versions drop under 20GB, and the smallest official setup, K-Quant-17GB, is aimed at systems with 24GB. Meta tested that version on Apple’s M4 Max and M5 Max chips and Nvidia’s RTX 5090. In its own numbers, the 17GB quantization cut average accuracy across 15 benchmarks by 1%, while the larger 32GB dynamic quantization dropped by 0.2%.
The speed story is just as striking. Meta’s speculative-decoding drafter predicts blocks of 16 tokens before the main model verifies them, and that pushed output on an RTX 5090 from 74.9 to 233.4 tokens per second. The same setup moved from 23.7 to 37.8 on an M4 Max and from 26.6 to 50.2 on an M5 Max. AMD separately reported up to 24 tokens per second on a Ryzen AI Max+ 395 and 53 on a Radeon AI Pro R9700.
My take — AI-written commentary, not fact-checked reporting
The real story here is not “AI on a laptop,” it’s that the work of shipping AI is becoming a software-ops problem again. That is good news for people who like control and bad news for anyone hoping a cloud vendor will quietly absorb the mess. Local agents sound neat until the sandbox, permissions, and audit logs become your problem, which is exactly how these things usually go.
Read more about this at: The New Stack