Meta 正在把具备较强自主执行能力的 AI Agent，从云端大模型压缩到可以在个人电脑上持续运行的 Muse Glimmer 30B。
部分指标超过 Qwen3.6-27B 和 Gemma4-31B。
通过 4-bit 量化，将模型压缩到约 17GB，目标是在 24GB 显存设备上运行。
采用 DFlash 推测解码，在 RTX 5090 上实现约 233 tokens/s，相比普通解码提速约 3.1 倍。
Meta 最近牛逼克拉斯！
Meta releases Muse Glimmer 30B, an open-weight agent model for consumer hardware
AISummary
Meta has released Muse Glimmer 30B, a 30B open-weight model aimed at agentic tasks on personal computers, according to a post by Dongxi NLP that quotes Meta's announcement.
The post says 4-bit quantization brings the model to about 17GB for 24GB GPUs, and that DFlash speculative decoding reaches about 233 tokens/s on an RTX 5090, roughly 3.1 times faster than standard decoding.
It also says some metrics exceed Qwen3.6-27B and Gemma4-31B.
Post on XView on X
Dongxi NLPVerified on X
@dongxi_nlp
Today we published a technical report on Muse Glimmer, our 30B open-weight model built for agentic tasks on consumer hardware. It runs at practical speeds on a Mac or PC with a single consumer GPU at 4-bit precision, and leads comparably-sized open models on tool use and multi-step planning. Read about the model architecture, distillation recipe, quantization and more: https://go.meta.me/4Bcvxol
Source: Dongxi NLP · x.comPublished · added here