Muse Glimmer評測：Meta開源30B代理模型單張GPU可跑 | Muse Glimmer Review: Meta's Open 30B AI Runs on One GPU

8月 13, 2026

By Kit 小克 | AI Tool Observer | 2026-08-13

🇹🇼 Muse Glimmer評測：Meta開源30B代理模型單張GPU可跑

Meta旗下Superintelligence Labs在8月10日發布Muse Glimmer，一款300億參數的開源代理模型，主打「單張消費級GPU就能跑」——只要24GB顯存（例如RTX 4090或3090），不必租雲端GPU、不用按token付費，就能在本機跑代理型任務。這是繼先前的程式碼代理Muse Code之後，Meta另一款瞄準「本機Agent」場景的開源模型。

Muse Glimmer是什麼？

Muse Glimmer是從Meta內部更大的閉源模型Spark 1.2蒸餾而來的300億參數密集模型，採用Apache 2.0授權，支援128K token上下文（可再延伸）。權重已上架Hugging Face，可透過Ollama、LM Studio、Unsloth直接下載本機跑，也能用llama.cpp、ExecuTorch、MLX部署到邊緣裝置，或用vLLM、SGLang大規模服務；Together AI、Fireworks AI、OpenRouter也已提供雲端API選項，不想自己架環境也能用。

跑分表現如何？

Meta官方公布的數據把Muse Glimmer拿去跟Gemma4-31B、Qwen3.6-27B比較，在DeepSearch QA、MCP-Atlas、τ3-Bench、SWE-Bench等「完整任務型」測試中表現不錯。這些測試主要衡量模型在多輪對話中呼叫工具、寫程式除錯、完成多步驟任務的能力，比單純問答評測更貼近真實Agent工作場景。不過跑分是廠商自己公布的，實際上手落差多大，還是得自己測。

對開發者的實際意義

省成本：本機跑代理任務不用按token計費，適合長時間掛機的排程、檔案整理類工作
隱私：敏感資料不必經過雲端API，適合企業內部或個人資料處理場景
門檻不是零：24GB顯存仍是入門門檻，一般筆電、舊顯卡跑不動；社群量化版（GGUF）能降低需求，但速度與品質會打折
官方定價未定：截至發稿，Meta尚未公布Muse Glimmer專屬的按token計費，第三方平台各自訂價，速度與費用要自己比較

小克實測心得

對比雲端旗艦模型，Muse Glimmer的重點不是「更聰明」，而是「夠用又能自己掌控」。如果你的工作流程是排程、檔案管理、內部工具串接這類重複性代理任務，值得抓下來試跑；但需要頂尖推理能力的複雜任務，還是建議搭配雲端模型使用。

好不好用，試了才知道。

🇺🇸 Muse Glimmer Review: Meta's Open 30B AI Runs on One GPU

Meta's Superintelligence Labs released Muse Glimmer on August 10 — a 30-billion-parameter open-weight agentic model built to run on a single consumer GPU with just 24GB of VRAM (think RTX 4090 or 3090). No cloud rental, no per-token API bill, just local agent tasks running on your own machine. It follows Meta's earlier coding agent Muse Code, this time targeting local, always-on agent workflows rather than coding specifically.

What Is Muse Glimmer?

Muse Glimmer is a dense 30B model distilled from Meta's larger closed model, Spark 1.2, released under the Apache 2.0 license with a 128K token context window (extendable further). Weights are live on Hugging Face — downloadable and runnable locally through Ollama, LM Studio, and Unsloth, deployable to edge devices via llama.cpp, ExecuTorch, and MLX, or served at scale with vLLM and SGLang. Together AI, Fireworks AI, and OpenRouter already list it for cloud-hosted access if you'd rather skip the local setup entirely.

How Does It Benchmark?

Meta's own numbers pit Muse Glimmer against Gemma4-31B and Qwen3.6-27B on full-task benchmarks like DeepSearch QA, MCP-Atlas, τ3-Bench, and SWE-Bench — tests that measure tool-calling, code debugging, and multi-turn task completion rather than simple Q&A, making them closer to real agent work. That said, these are vendor-reported figures, so it's worth verifying real-world performance yourself before betting a workflow on it.

What It Actually Means for Developers

Lower cost: no per-token billing for always-on scheduling, file management, or background agent tasks
Privacy: sensitive data never leaves your machine — useful for internal tools or personal data processing
The hardware floor isn't zero: 24GB VRAM is still required; most laptops and older GPUs can't run it natively, and community GGUF quantizations lower the bar at the cost of speed and quality
No official Glimmer-specific pricing yet: as of publication, Meta hasn't announced per-token API pricing — third-party hosts set their own rates and speeds

Kit's Honest Take

The pitch here isn't that Muse Glimmer beats frontier cloud models on raw intelligence — it's that you get "good enough, fully under your control." If your workload is repetitive local agent tasks like scheduling or file organization, it's worth pulling down and testing. For anything needing top-tier reasoning on complex tasks, pairing it with a cloud model still makes sense.

好不好用，試了才知道 — good or not, you won't know until you try it.

Sources / 資料來源

延伸閱讀 / Related Articles

AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends

搜尋此網誌

AI小貼士