Muse Glimmer: 30B Open Agentic Model for Local Agent Workflows
Source: https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model
Author: Meta Superintelligence Labs
Date: 2026-08-10
Category: models
Type: article
Quality: 5/5
Tags: meta, open-weights, local-agent, on-device, 30b, apache-2
English Summary
Meta Superintelligence Labs released Muse Glimmer, a 30-billion-parameter open-weight model (Apache 2.0) optimized for always-on local agent workflows. It runs on a single consumer GPU (fitting under 20GB via 4-bit quantization) and supports function calling, local coding, multimodal input, and LLM-as-a-judge evaluation. Trained via logit distillation from Muse Spark, mid-training on agent-heavy data, and post-training combining SFT with on-policy distillation and RL. Ships with DFlash speculative decoding for 1.5-3.1x speedup. Evaluated on DeepSearch QA, MCP-Atlas, tau-Bench, and SWE-Bench. Available on Hugging Face with llama.cpp, MLX, ExecuTorch integrations.
中文概要
Meta 超级智能实验室发布 Muse Glimmer，一个 300 亿参数的开源模型 (Apache 2.0)，专为本地常驻代理工作流优化通过 4-bit 量化压缩至 20GB 以下，可在单块消费级 GPU 上运行，支持函数调用本地编码多模态输入和 LLM-as-a-judge 评估训练经历 logit 蒸馏代理密集型中间训练以及 SFT+在策略蒸馏+RL 后训练，并随附 DFlash 推测解码加速 1.5-3.1 倍在 DeepSearch QAMCP-Atlas-BenchSWE-Bench 等基准上评测，权重已在 Hugging Face 开放
*Added via external-scan on 2026-08-11*