Meta Open‑Sources Muse Glimmer: Packing a 30B Model into 24 GB VRAM for Always‑On Local Agents

On August 10, Meta Superintelligence Labs announced Muse Glimmer, an open‑source 30‑billion‑parameter dense model (≈29.6 B parameters) released under the Apache 2.0 license. The model is explicitly optimized for “always‑on local agent workflows,” targeting personal computers equipped with a single consumer‑grade GPU.

The architecture combines a causal Transformer backbone with a 1.8 B‑parameter visual perception encoder, supporting both text and image inputs and a context window exceeding 131 k tokens. Rather than pursuing ever‑larger parameter counts, the design prioritizes retaining long‑context, multimodal, and agentic abilities within the memory limits of consumer hardware.

Training follows a three‑stage pipeline. First, a pre‑training phase uses logit distillation from the Muse Spark model. The mid‑training stage raises the proportion of long‑context data, agent‑centric examples, and reasoning traces. Finally, post‑training blends supervised fine‑tuning, on‑policy distillation, and reinforcement learning to improve general reasoning, coding, and agentic tasks (Meta AI Research).

To fit the model into 24–32 GB VRAM, Meta applies roughly 4‑bit quantization, shrinking the weight footprint to under 20 GB. Two quantization targets are provided: K‑Quant‑Dynamic for 32 GB VRAM (average benchmark accuracy drop ≈0.2 %) and K‑Quant‑17GB for 24 GB VRAM (average drop ≈1.0 %). This compression leaves room for KV‑cache, the visual encoder, and a speculative decoding drafter.

Benchmark results show Muse Glimmer excelling on several agentic tasks. On the MCP Atlas benchmark it scores 75.5, compared with 54.2 for Gemma‑4‑31B and 62.5 for Qwen‑3.6‑27B. DeepSearch QA yields 74.6 versus 61.7 and 71.1, and SWE‑Bench Pro records 51.2 versus 36.9 and 50.2. The model does not dominate every test—Qwen‑3.6‑27B outperforms it on SWE‑Bench Verified and OSWorld‑Verified—highlighting its deliberate trade‑off between size, deployment cost, and agentic performance.

Speed is addressed with a DFlash speculative decoding drafter that predicts 16‑token blocks for the main model to verify. On an RTX 5090, standard decoding reaches 74.9 tok/s, while DFlash raises average throughput to 233.4 tok/s (≈3.1×). On Apple silicon, M5 Max improves from 26.6 tok/s to 50.2 tok/s (≈1.8×) and M4 Max from 23.7 tok/s to 37.8 tok/s (≈1.5×). This 233 tok/s figure is presented as the threshold for practical, continuous agent operation.

The authors argue that the significance of Muse Glimmer lies not in adding another open‑source model to a leaderboard, but in redefining the deployment boundary for personal‑device agents. By keeping the model resident in memory and achieving near‑real‑time inference, a hybrid architecture becomes feasible: the local model handles high‑frequency, low‑latency tasks (file reading, screenshot understanding, tool calling, coding, failure recovery), while a larger cloud model is invoked only for complex reasoning.

In summary, Muse Glimmer demonstrates that a 30 B‑parameter, multimodal, long‑context model can be quantized, accelerated, and run continuously on consumer hardware, thereby shifting the paradigm from “cloud‑only agents” to “always‑on local agent runtimes.”