Muse Glimmer評測：Meta回歸開源，30B模型單張顯卡就能跑 | Muse Glimmer Review: Meta's 30B Open Model Runs on One GPU

8月 27, 2026

By Kit 小克 | AI Tool Observer | 2026-08-28

🇹🇼 Muse Glimmer評測：Meta回歸開源，30B模型單張顯卡就能跑

Muse Glimmer 是 Meta 睽違一年多後重新推出的開源模型，300 億參數、Apache 2.0 授權，重點不是刷榜，而是壓縮到單張消費級顯卡就能在家跑起來。這篇評測整理它的規格、實測效能，以及為什麼 Meta 這時候選擇回頭擁抱開源。

Muse Glimmer規格：300億參數、4-bit量化下不到20GB

Muse Glimmer 於 2026 年 8 月 10 日發布，訓練資料源自 Meta 自家的 Muse Spark 系列模型，再經過兩輪針對長文推理與 Agent 任務的強化。重點規格：

參數量：300 億（30B）
授權：Apache 2.0，商用不受限
硬體門檻：4-bit 量化後壓縮到 20GB 以內，MacBook M4-Max / M5-Max 或單張 RTX 5090（24–32GB 記憶體）就能跑
加速技術：搭配 DFlash 推測解碼（speculative decoding），RTX 5090 上實測加速 3.1 倍，M5-Max 加速 1.8 倍
下載管道：權重已上架 Hugging Face，llama.cpp、MLX、ExecuTorch 的整合近日陸續到位

效能實測：打平Qwen3.6與Gemma4，但不是全面領先

Meta 拿 Muse Glimmer 與同量級的 Gemma4-31B、Qwen3.6-27B 在近二十個基準測試上比較，結果是約半數項目勝出，強項集中在線上檢索、程式碼生成、圖表理解等 Agent 類任務，多模態輸入（文字+圖片）與超過 100 種語言也在支援範圍。換句話說，Muse Glimmer 不是同噸位裡的全能冠軍，但作為「能在自己電腦上跑的開源模型」，這個效能水準已經算誠意十足。

為什麼Meta現在回頭做開源

過去一年 Meta 在開源策略上明顯轉向保守，Llama 4 系列一度傳出縮減開放範圍。這次 Muse Glimmer 上線，執行長 Zuckerberg 同步發表文章談 AI 風險與治理，並承諾 Muse Spark 1.2 的完整權重「即將」開放——時間點也正好卡在 Qwen、DeepSeek、GLM 等中國開源模型持續放大招之際。對開發者來說，這代表本地部署開源模型的選項又多了一個能打的。

好不好用，試了才知道。

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

Muse Glimmer is Meta's first open-weight model release in over a year — 30 billion parameters, Apache 2.0 licensed, and built specifically to run on a single consumer GPU rather than chase leaderboard scores. Here is what it actually ships with, how it performs, and why Meta picked this moment to return to open source.

Muse Glimmer Specs: 30B Params, Under 20GB at 4-bit

Released August 10, 2026, Muse Glimmer is trained on data from Meta's proprietary Muse Spark models, with two additional refinement passes for long-context reasoning and agentic tasks. Key specs:

Size: 30 billion parameters
License: Apache 2.0, unrestricted commercial use
Hardware: compresses to under 20GB at 4-bit quantization — runs on a MacBook M4-Max/M5-Max or a single RTX 5090 (24–32GB memory)
Speed: paired with a DFlash speculative decoding drafter, measured at 3.1x speedup on RTX 5090, 1.8x on M5-Max
Availability: weights are live on Hugging Face; llama.cpp, MLX, and ExecuTorch integrations are rolling out within days

Benchmarks: Trades Blows with Qwen3.6 and Gemma4

Meta benchmarked Muse Glimmer against similarly-sized Gemma4-31B and Qwen3.6-27B across roughly two dozen tests, and it won about half of them — strongest in online research, code generation, and chart/image understanding for agentic workflows. It also handles multimodal input (text and images) and over 100 languages. It's not a category-leading model outright, but as an open weight you can actually run at home, the performance is a genuinely honest showing.

Why Meta Is Returning to Open Source Now

Meta had been quietly walking back its open-source commitments over the past year, with reports of narrower releases in the Llama 4 line. Muse Glimmer's launch came alongside a Zuckerberg essay on AI risk and governance, plus a promise that full weights for Muse Spark 1.2 are coming soon. The timing lines up with intensifying pressure from Chinese open-weight labs — Qwen, DeepSeek, GLM — all shipping aggressively this year. For developers, it means one more serious option for running open models locally.

You won't know if it's good 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