Meta Muse Spark 模型完整解析:Alexandr Wang 操刀的閉源轉向,跑分第四但效率提升 10 倍 | Meta Muse Spark Explained: Alexandr Wang Ditches Open-Source, Ranks 4th But 10x More Efficient
By Kit 小克 | AI Tool Observer | 2026-04-12
🇹🇼 Meta Muse Spark 模型完整解析:Alexandr Wang 操刀的閉源轉向,跑分第四但效率提升 10 倍
Meta Muse Spark 是 Meta 在 2026 年 4 月 8 日發布的全新 AI 模型,也是 Meta 超智慧實驗室(Meta Superintelligence Labs)成立以來的第一個作品。由前 Scale AI 創辦人 Alexandr Wang 擔任首席 AI 長親自操刀,這個代號「Avocado」的模型花了九個月打造,直接取代了 Llama 4 系列的定位。
Meta Muse Spark 是什麼?跟 Llama 有什麼不同?
Muse Spark 是 Meta 全新 Muse 系列的首發模型,最大的改變是閉源策略。Meta 過去以 Llama 系列的開源路線聞名,但 Muse Spark 走的是專有路線,官方僅表示「希望未來能開源部分版本」。這個轉向引發社群不小的爭議。
Muse Spark 跑分表現如何?值得用嗎?
根據 Artificial Analysis Intelligence Index v4.0,Muse Spark 綜合排名第四(52 分),落後 Gemini 3.1 Pro(57)、GPT-5.4(57)和 Claude Opus 4.6(53)。但在特定領域表現突出:
- 醫療領域:HealthBench Hard 拿下 42.8%,擊敗所有競爭對手
- 視覺理解:MMMU-Pro 80.5%,僅次於 Gemini 3.1 Pro 的 82.4%
- 深度推理:Contemplating 模式在 Humanity’s Last Exam 達 58%
Muse Spark 的三種推理模式怎麼選?
Muse Spark 提供三種推理模式,適合不同場景:
- Instant:日常問答,快速回應,不做深度推理
- Thinking:延伸思維鏈,適合較難的問題
- Contemplating:多 Agent 並行推理,對標 Gemini Deep Think 和 GPT Pro 的極限推理模式
效率提升才是真正的亮點
Meta 官方表示,Muse Spark 達到相同能力所需的算力比 Llama 4 Maverick 少超過一個數量級。這代表 Meta 在模型架構上做了根本性的改進,不只是堆更多 GPU。對開發者來說,這意味著更低的推理成本。
免費使用,但有限制
目前 Muse Spark 可在 meta.ai、Meta AI App、以及 Facebook、Instagram、WhatsApp、Messenger 中免費使用。也支援 Ray-Ban Meta AI 智慧眼鏡。不過目前僅支援文字輸出,接受語音、文字和圖片輸入。
整體來說,Muse Spark 不是最強的模型,但考慮到效率提升和免費策略,它在 Meta 生態系內的實用性不容小覷。只是從開源走向閉源這步棋,讓不少開發者感到失望。好不好用,試了才知道。
🇺🇸 Meta Muse Spark Explained: Alexandr Wang Ditches Open-Source, Ranks 4th But 10x More Efficient
Meta Muse Spark launched on April 8, 2026, as the first model from Meta Superintelligence Labs. Built over nine months by a team led by chief AI officer Alexandr Wang (formerly Scale AI co-founder), this model—codenamed Avocado—represents a fundamental shift in Meta’s AI strategy.
What Is Muse Spark and How Does It Differ from Llama?
Muse Spark is the debut model in Meta’s new Muse series, and the biggest change is its proprietary, closed-source approach. Meta previously championed open-source AI through its Llama series, but Muse Spark breaks that tradition. The company says it “hopes to open-source future versions,” but the community reaction has been mixed.
How Does Muse Spark Perform on Benchmarks?
On the Artificial Analysis Intelligence Index v4.0, Muse Spark ranks 4th overall with a score of 52, behind Gemini 3.1 Pro (57), GPT-5.4 (57), and Claude Opus 4.6 (53). However, it excels in specific domains:
- Healthcare: 42.8% on HealthBench Hard, beating all competitors
- Vision: 80.5% on MMMU-Pro, second only to Gemini 3.1 Pro (82.4%)
- Deep reasoning: Contemplating mode hits 58% on Humanity’s Last Exam
Three Reasoning Modes for Different Use Cases
Muse Spark offers three reasoning modes: Instant for quick, casual queries; Thinking for extended chain-of-thought reasoning on harder problems; and Contemplating, which orchestrates multiple agents reasoning in parallel—competing directly with Gemini Deep Think and GPT Pro.
The Real Story: 10x Efficiency Gains
Meta claims Muse Spark achieves the same capabilities with over an order of magnitude less compute than Llama 4 Maverick. This suggests fundamental architectural improvements, not just more GPUs. For developers, this translates to significantly lower inference costs.
Free to Use, But With Caveats
Muse Spark is available for free on meta.ai, the Meta AI app, and across Facebook, Instagram, WhatsApp, Messenger, and Ray-Ban Meta AI glasses. It accepts voice, text, and image inputs but currently outputs text only.
Muse Spark isn’t the strongest model on the market, but its efficiency gains and zero-cost access make it a compelling option within Meta’s ecosystem. The shift from open-source to closed-source, though, has left many developers disappointed. As always—you won’t know until you try it yourself.
Sources / 資料來源
- Meta debuts Muse Spark, first AI model under Alexandr Wang - Axios
- Meta debuts new AI model after spending billions - CNBC
- Muse Spark: Meta is back in the AI race - Artificial Analysis
常見問題 FAQ
Meta Muse Spark 是免費的嗎?
是的,Muse Spark 目前免費使用,可在 meta.ai、Meta AI App 及 Facebook、Instagram、WhatsApp 中使用,但可能有速率限制。
Muse Spark 跟 Llama 有什麼不同?
Muse Spark 是閉源模型,屬於全新的 Muse 系列,由 Alexandr Wang 領導的 Meta 超智慧實驗室開發,取代了之前的開源 Llama 路線。
Muse Spark 的推理能力排第幾?
在 Artificial Analysis Intelligence Index v4.0 中排名第四(52分),落後 Gemini 3.1 Pro、GPT-5.4 和 Claude Opus 4.6。
Muse Spark 最強的領域是什麼?
醫療領域,在 HealthBench Hard 測試中以 42.8% 擊敗所有競爭對手,包括 GPT-5.4 和 Claude Opus 4.6。
延伸閱讀 / Related Articles
- Vitalik Buterin 本地 LLM 方案實測解析:NVIDIA 5090 跑 Qwen3.5,不上雲也能日常用 AI | Vitalik Buterin Local LLM Setup Explained: Running Qwen3.5 on NVIDIA 5090 — Daily AI Without the Cloud
- Google TurboQuant 實測解析:KV Cache 壓縮到 3-bit,記憶體砍 6 倍還不掉精度 | Google TurboQuant Explained: 3-bit KV Cache Compression Cuts Memory 6x With Zero Accuracy Loss
- Google AI Edge Gallery 實測:手機離線跑 Gemma 4,AI Agent 不用網路也能動 | Google AI Edge Gallery Hands-On: Running Gemma 4 Offline on Your Phone — AI Agents Without Internet
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
張貼留言