FULL STORY

Meta's Muse Glimmer: A 30B On-Device Agent Model

Meta released Muse Glimmer, a 30B open-weight model for on-device agent tasks. Engineers followed up with details of how a small team trained it from scratch and fit it into 32GB of VRAM.

2026-10-20 ~ 2026-10-21 · 2 episodes · 10 posts

Episode 1 · Meta Open-Sources Muse Glimmer 30B, an On-Device Agent Model (2026-10-20, 5 posts)

Meta has released Muse Glimmer, a 30-billion-parameter open-weights model designed to run agentic tasks locally on consumer hardware. It comes out of Meta Superintelligence Labs and is open-sourced under Apache 2.0.

Confirmed

  • The model is distilled from Meta's frontier model Muse Spark and purpose-built for autonomous Agent tasks on consumer hardware
  • With 4-bit quantization, it achieves practical speeds on a Mac or PC with a single 32GB consumer GPU
  • A technical report has been released; it performs well on agentic benchmarks such as tool calling and multi-step planning (compared with open models of similar size)
  • @TimDarcet noted the model was unveiled at Connect, presented by FAIR research engineering manager Maryam Fazel-Zarandi and researcher Yoram

Why it matters

  • This is Meta's open-source play for the "local agent" use case, distilling frontier-model capabilities down to a size that runs on personal devices and lowering the hardware barrier for deploying agents locally
  • Open weights under Apache 2.0 mean developers can freely use it commercially and build on it, potentially boosting the on-device Agent ecosystem

Episode 2 · Meta team trains 30B LLM from scratch in months, squeezes it into a 32GB GPU (2026-10-21, 5 posts)

Meta engineer TimDarcet shared that he and colleagues including Parth, Kelvin, and Anton spent a few months building a team and a complete training pipeline, training an LLM called Muse Glimmer-30B from scratch. He described the results as "quite good" and sees it as a promising starting point for a "second era" of Meta open source. Although the model fits in VRAM, memory bandwidth and throughput remain bottlenecks, so the team tackled deployment via two routes: quantization and speculative decoding.

Confirmed

  • The model is Muse Glimmer-30B, trained from scratch by a small team within a few months.
  • The EvoPress quantization scheme was obtained via evolutionary search: average accuracy 75.6% (vs. 75.7% for bf16), average response length 11417 (vs. 11070 for bf16), with weight size reduced from 56GB to 19GB—nearly lossless.
  • The quantization scheme was found automatically through greedy and evolutionary search, jointly minimizing VRAM usage and perplexity error, yielding two Pareto-optimal trade-off points; TimDarcet confirmed the results are Pareto-optimal.
  • Equipped with a DFlash speculative decoding drafter: 5-layer draft, block size 16 (predicting 16 tokens at a time), to relieve memory bandwidth and throughput pressure.
  • The end result: a 30B-parameter model that runs on a 32GB GPU.

Why it matters

This work demonstrates a small team completing a full-chain practice—from data to training, quantization, and inference acceleration—within a few months. In particular, the combination of search-based quantization and speculative decoding offers a referenceable path for deploying mid-scale models on consumer GPUs. giffmana also relayed the quantization research approach and jokingly asked why agentic search wasn't used, indirectly reflecting how much attention this direction is attracting.