How Meta’s Open‑Source 30B Muse Glimmer Agent Runs on Your PC
Meta’s newly open‑sourced 30‑billion‑parameter Muse Glimmer agent model runs on a single consumer‑grade GPU, outperforms Gemma‑4 and Qwen‑3.6 on multiple Agent benchmarks, uses a perception encoder for multimodal input, and fits into a 20 GB memory envelope through quantization and a lightweight drafter.
What It Can Do
Muse Glimmer is a 30 B parameter agent that can run on a personal Mac or PC with a single consumer‑grade GPU. It leads several Agent benchmarks—MCP‑Atlas (75.5), DeepSearch QA (74.6), τ³‑Banking (23.5), WildClawBench (47.6), GAIA2 (43.3)—and also tops coding (SWE‑Bench Pro 51.2, SciCode 43.6), long‑context (AIME 2026 94.7, AA‑LCR 80.0, Beam 128K 65.1) and multimodal (Charxiv Reasoning 78.8) evaluations. The model handles function calling with precise schemas, diagnoses and retries failed tool calls, and maintains coherent multi‑step reasoning over long workflows.
Born for Local Use
The model is deliberately optimized for on‑device agents that manage schedules, draft messages, organize files, and learn personal work habits. Meta balances capability with the memory and compute limits of consumer hardware by using a compact architecture, a novel distillation recipe, and inference‑time quantization.
Fitting into 20 GB
Full‑precision 300 B‑parameter models require >55 GB of memory, far beyond consumer GPUs. Muse Glimmer’s weights are quantized to ~4‑bit, shrinking the model below 20 GB. The remaining memory holds the KV cache, a perception encoder for image‑text inputs, and a speculative‑decoding drafter based on DFlash, all fitting within a 24‑32 GB envelope. Meta reports negligible performance loss on Agent tasks.
The lightweight drafter proposes whole token blocks that the main model validates, yielding much faster generation than token‑by‑token decoding while preserving output quality.
Get Started in Minutes
The weights are hosted on Hugging Face with developer documentation. Integration with llama.cpp, MLX, and ExecuTorch is forthcoming. After downloading, an agent can be up and running in a few minutes. Users can run locally via Ollama, LM Studio, or Unsloth, scale with vLLM or SGLang, or customize further with PyTorch’s TorchTitan training.
Meta also collaborates with AMD, Arm, Dell, Intel, and NVIDIA to optimize performance across devices.
All resources, including model weights, benchmarks, and deployment guides, are available through Meta’s AI Developer Center.
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
SuanNi
A community for AI developers that aggregates large-model development services, models, and compute power.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.