Meta Muse Spark 模型解析：Alexandr Wang 操刀的首款閉源 AI，免費使用但放棄開源路線 | Meta Muse Spark Explained: First Closed-Source Model Under Alexandr Wang — Free to Use but Goodbye Open Source
By Kit 小克 | AI Tool Observer | 2026-04-21
🇹🇼 Meta Muse Spark 模型解析：Alexandr Wang 操刀的首款閉源 AI，免費使用但放棄開源路線
Meta 在 2026 年 4 月 8 日發布了 Muse Spark，這是 Meta 超智慧實驗室（Superintelligence Labs）在 Alexandr Wang 領導下推出的第一款 AI 模型。最大的爭議點？Meta Muse Spark 是閉源的，正式告別了 Llama 系列引以為傲的開源策略。
Meta Muse Spark 是什麼？有什麼能力？
Muse Spark 是一款原生多模態推理模型，支援語音、文字和圖片輸入，目前僅輸出文字。它有快速模式處理日常問答，也有多種推理模式應對複雜問題。最值得注意的是，Muse Spark 的 token 效率極高——完成完整的 Intelligence Index 評測只需 5800 萬個輸出 token，遠少於 Claude Opus 4.6 的 1.57 億和 GPT-5.4 的 1.2 億。
Meta Muse Spark 跑分表現如何？
Muse Spark 在 Artificial Analysis Intelligence Index 排名第四，得分 52，落後於 Gemini 3.1 Pro、GPT-5.4 和 Claude Opus 4.6。具體表現：
- 強項：圖表理解（86.4）、HealthBench Hard（42.8%），多模態理解是真正的亮點
- 弱項：Terminal-Bench 2.0 僅 59.0（GPT-5.4 達 75.1）、ARC AGI 2 思考模式 42.5（對手 76+）
簡單說：多模態很強，但程式碼和抽象推理還有明顯差距。
為什麼 Meta Muse Spark 放棄開源？
這是最讓社群震驚的決定。Meta 過去靠 Llama 開源建立了龐大的開發者生態系，但 Muse Spark 選擇閉源，優先考慮競爭優勢而非社群採用。Meta 表示「未來會釋出某個版本」的開源授權，但沒有給出時程。目前 Meta 採取雙軌策略：Llama 繼續負責社群生態，Muse 系列專注競爭產品。
Meta Muse Spark 怎麼收費？誰能用？
目前 Meta Muse Spark 完全免費，透過 Meta AI app 和 meta.ai 網站即可使用（可能有速率限制）。API 存取目前僅限受邀夥伴私測，公開付費 API 尚未公布定價。Meta 的商業模式不靠 AI 服務收費，而是透過「購物模式」整合 Instagram 和 Threads 的用戶行為數據，推薦個人化商品，用 30 億用戶的廣告和電商收入來支撐 AI 成本。
Alexandr Wang 帶來了什麼改變？
2025 年 6 月，Meta 花了 143 億美元收購 Scale AI 49% 股份，並把共同創辦人 Alexandr Wang 挖來擔任首任首席 AI 長。Wang 的團隊花了九個月從頭重建 AI 基礎設施。他說：「九個月前我們從零重建了 AI 技術棧——新基礎設施、新架構、新數據管線。」Meta 2026 年的 AI 資本支出預算高達 1150 至 1350 億美元，幾乎是去年的兩倍。
Kit 小克的觀察
Meta Muse Spark 目前排第四，坦白說還追不上第一梯隊。但免費使用加上 30 億用戶基礎，這個分發優勢是其他 AI 公司夢寐以求的。閉源決定雖然讓開源社群失望，但從商業角度看完全合理——你不會把最好的武器送給對手。值得關注的是即將推出的「沉思模式」（Contemplating Mode），用多個子 Agent 平行推理，這可能是追趕 GPT-5.4 Pro 和 Gemini Deep Think 的關鍵。
好不好用，試了才知道。
常見問題 FAQ
Meta Muse Spark 可以免費使用嗎？
可以，目前透過 Meta AI app 和 meta.ai 網站完全免費使用，可能有速率限制。公開付費 API 尚未推出。
Meta Muse Spark 比 ChatGPT 好嗎？
在多模態理解和醫療健康任務上，Muse Spark 表現出色。但在程式碼生成和抽象推理方面，目前仍落後於 GPT-5.4 和 Claude Opus 4.6。
Meta 以後還會做開源模型嗎？
會，Llama 系列仍會繼續開源更新。但 Meta 最強的模型將走閉源路線，開源社群能拿到的版本可能會有延遲或功能縮減。
Muse Spark 的 token 效率為什麼重要？
更少的 token 意味著更快的回應速度和更低的運算成本。Muse Spark 只用競爭對手三分之一到一半的 token 就完成同樣的評測，這對大規模部署特別有利。
🇺🇸 Meta Muse Spark Explained: First Closed-Source Model Under Alexandr Wang — Free to Use but Goodbye Open Source
On April 8, 2026, Meta released Muse Spark — the first AI model from Meta Superintelligence Labs under the leadership of Alexandr Wang. The biggest controversy? Meta Muse Spark is closed-source, marking a formal departure from the open-source strategy that defined the Llama era.
What Is Meta Muse Spark and What Can It Do?
Muse Spark is a natively multimodal reasoning model that accepts voice, text, and image inputs with text-only output. It features a fast mode for casual queries and multiple reasoning modes for complex problems. Most notably, Muse Spark is remarkably token-efficient — it completed the full Intelligence Index evaluation using just 58 million output tokens, compared to 157 million for Claude Opus 4.6 and 120 million for GPT-5.4.
How Does Meta Muse Spark Perform on Benchmarks?
Muse Spark ranks 4th on the Artificial Analysis Intelligence Index with a score of 52, trailing Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6. Specifics:
- Strengths: Figure understanding (86.4), HealthBench Hard (42.8%) — multimodal perception is the real highlight
- Weaknesses: Terminal-Bench 2.0 at 59.0 (vs GPT-5.4 at 75.1), ARC AGI 2 thinking at 42.5 (vs 76+ for rivals)
Bottom line: excellent at multimodal tasks, but coding and abstract reasoning still lag behind.
Why Did Meta Muse Spark Go Closed-Source?
This is the most surprising decision. Meta built an enormous developer ecosystem through open-source Llama models, but Muse Spark prioritizes competitive advantage over community adoption. Meta says it will release "a version" under open-source licensing eventually, but has given no timeline. The current approach is dual-track: Llama for community, Muse for competitive products.
How Much Does Meta Muse Spark Cost?
Meta Muse Spark is currently free via the Meta AI app and meta.ai website, with possible rate limits. API access is limited to invited partners; public paid API pricing has not been announced. Meta monetizes through enhanced advertising and commerce — a "shopping mode" integrates Instagram and Threads activity for personalized product recommendations, leveraging its 3-billion-user base rather than charging per token.
What Has Alexandr Wang Changed at Meta?
In June 2025, Meta spent $14.3 billion acquiring a 49% stake in Scale AI and hired co-founder Alexandr Wang as its first Chief AI Officer. Wang and his team spent nine months rebuilding Meta's AI stack from scratch. As Wang said: "Nine months ago we rebuilt our AI stack from scratch — new infrastructure, new architecture, new data pipelines." Meta's 2026 AI capital expenditure budget is $115–135 billion, nearly double last year.
Kit's Take
Muse Spark currently ranks 4th — honestly, it hasn't caught the top tier yet. But free access combined with a 3-billion-user distribution base gives Meta an advantage other AI companies can only dream of. The closed-source decision disappointed the open-source community, but it makes perfect business sense — you don't hand your best weapon to competitors. Watch for the upcoming "Contemplating Mode" that deploys multiple sub-agents reasoning in parallel — this could be Meta's key to catching GPT-5.4 Pro and Gemini Deep Think.
You never know until you try.
FAQ
Is Meta Muse Spark free to use?
Yes, it is currently free through the Meta AI app and meta.ai website, with possible rate limits. A public paid API has not launched yet.
Is Meta Muse Spark better than ChatGPT?
Muse Spark excels at multimodal understanding and healthcare tasks. However, it still trails GPT-5.4 and Claude Opus 4.6 in code generation and abstract reasoning.
Will Meta still release open-source models?
Yes, the Llama series will continue with open-source updates. But Meta's most powerful models will be closed-source, and open-source versions may come with delays or reduced capabilities.
Why does Muse Spark's token efficiency matter?
Fewer tokens mean faster responses and lower compute costs. Muse Spark uses one-third to one-half the tokens of competitors for equivalent evaluations, which is a major advantage for large-scale deployment.
Sources / 資料來源
- TechCrunch: Meta debuts the Muse Spark model
- Axios: Meta debuts Muse Spark under Alexandr Wang
- danilchenko.dev: Muse Spark ditches Llama open-source playbook
常見問題 FAQ
Meta Muse Spark 可以免費使用嗎？
可以，目前透過 Meta AI app 和 meta.ai 網站完全免費，公開付費 API 尚未推出。
Meta Muse Spark 比 ChatGPT 好嗎？
多模態和醫療任務表現出色，但程式碼和抽象推理仍落後 GPT-5.4 和 Claude Opus 4.6。
Meta 以後還會做開源模型嗎？
Llama 系列繼續開源，但最強模型走閉源路線，開源版本可能延遲或功能縮減。
Muse Spark 的 token 效率為什麼重要？
更少 token 代表更快回應和更低成本，Muse Spark 只用競爭對手三分之一到一半的 token。
延伸閱讀 / Related Articles
- Microsoft Agent Governance Toolkit 開源解析：首款覆蓋 OWASP 十大風險的 AI Agent 安全框架 | Microsoft Agent Governance Toolkit Explained: First Open-Source Framework Covering All 10 OWASP Agentic AI Risks
- Claude Mythos 5 十兆參數模型解析：MoE 架構只啟動 1/10 參數，資安與學術推理能力遠超 Opus 4.6 | Claude Mythos 5 Explained: 10 Trillion Parameters With MoE Activating Only 1/10 — Cybersecurity and Research Reasoning Leap Beyond Opus 4.6
- NVIDIA Ising 量子 AI 模型解析：35B 參數校準 VLM + 即時糾錯 CNN，量子電腦終於有 AI 作業系統了 | NVIDIA Ising Quantum AI Models Explained: 35B Calibration VLM + Real-Time Error Correction CNN — Quantum Computing Finally Gets an AI Operating System
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