跳到主要內容

Muse Spark Open評測:Meta開源816B模型,Mac可跑 | Muse Spark Open Review: Meta's 816B Model Fits on Mac

By Kit 小克 | AI Tool Observer | 2026-10-16

🇹🇼 Muse Spark Open評測:Meta開源816B模型,Mac可跑

Meta在2026年10月15日開源了Muse Spark Open,一款816B參數、42B啟動參數的混合專家(MoE)模型,官方稱它是目前美國發布過最強的開源權重模型。這篇用實測角度拆解它到底多能打,以及一般人有沒有機會自己跑起來。

Muse Spark Open規格重點

  • 總參數816B,但每次推理只啟動約42B,這就是為什麼號稱能在單台Mac Studio上跑
  • 原生多模態:文字、圖片、音訊、影片都吃
  • context window衝到100萬token
  • 主打coding與agentic任務,不是純聊天模型
  • 支援vLLM、SGLang推理引擎,NVIDIA/AMD/Intel都有優化

老實說:Mac Studio能跑是什麼意思

MoE架構的好處是模型總體積816B看起來嚇人,但實際運算只動用42B參數,記憶體頻寬需求因此大幅下降。這也是Muse Spark Open被媒體拿來講「可在Mac Studio跑」的關鍵——但別誤會,不是隨便一台筆電就能開,你還是需要頂規Mac Studio(通常是512GB統一記憶體版本)搭配量化版模型才跑得動,一般消費級顯卡幾乎沒戲。

跟GPT、Gemini比起來呢

Meta發布時沒有附上對比Claude或GPT-6等級的跑分,第三方測試還在陸續公布。老實說開源權重模型目前最大價值不是「贏過誰」,而是你能自己架、自己改、資料不用出門。如果你的需求用雲端API就能解決,Muse Spark Open不會讓你更快更省;但如果你要做本地agent、要客製化微調,這是目前能拿到手最大的開源多模態模型。

授權與安全性

這個開源權重版本經過額外的safety post-training,目的是在沒有系統層guardrail保護時也不要太容易被濫用。權重可以從Hugging Face、Amazon SageMaker、Google Cloud Model Garden、Microsoft Foundry下載,也能直接用Baseten、Fireworks、Groq、Together AI等平台的API試跑,不用自己扛機器。

好不好用,試了才知道。


🇺🇸 Muse Spark Open Review: Meta's 816B Model Fits on Mac

Muse Spark Open is Meta's new open-weight model released on October 15, 2026 — an 816B-parameter mixture-of-experts model with only 42B active parameters per token, which Meta is calling the most capable open-weight model released in the US so far. Here's what actually matters if you're thinking about running it.

The Specs That Matter

  • 816B total parameters, ~42B active per forward pass — the MoE design is why this is runnable at all outside a data center
  • Natively multimodal: text, image, audio, and video in one model
  • 1M-token context window
  • Built specifically for coding and agentic workloads, not general chat
  • Runs on vLLM and SGLang, with optimization for NVIDIA, AMD, and Intel hardware

"Runs on a Mac Studio" — What That Actually Means

Headlines saying Muse Spark Open "runs on a Mac Studio" are technically true but easy to overread. The MoE architecture means only 42B of the 816B total parameters activate per token, which is why memory bandwidth requirements are manageable — but you still need a maxed-out Mac Studio (think the 512GB unified memory config) running a quantized version, not a base MacBook. Consumer GPUs are mostly out of the picture entirely.

How Does It Compare to GPT-6 or Gemini?

Meta didn't publish head-to-head benchmarks against Claude or GPT-6 at launch, and independent evaluations are still trickling in. Honestly, the real value of an open-weight model like Muse Spark Open isn't "beats everyone" — it's that you can self-host, fine-tune, and keep your data in-house. If a hosted API already solves your problem, this won't make things cheaper or faster. But if you need a local agent stack or custom fine-tuning, this is the largest open multimodal model you can actually get your hands on right now.

Licensing and Safety

The open-weight release went through additional safety post-training specifically because it won't have system-level guardrails wrapped around it the way the hosted version does. Weights are available on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models; or you can just hit the API on Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, or Together AI without running any hardware yourself.

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

留言

這個網誌中的熱門文章

Google Ironwood TPU v7 推理專用晶片解析:效能追平 NVIDIA、成本低 44%,AI 晶片戰爭正式開打 | Google Ironwood TPU v7 Explained: Matching NVIDIA Performance at 44% Lower Cost — The AI Chip War Heats Up

Claude Code 實測:AI 幫你寫程式到底行不行? | Claude Code Review: Can AI Really Code for You?

Cursor vs GitHub Copilot vs Claude Code:AI 程式助手大比拼 | AI Coding Assistants Compared: Cursor vs GitHub Copilot vs Claude Code