Muse Spark Open 對決 Muse Wonder：一個可下載，一個只能租用

Meta put Muse Wonder on the market on October 12, 2026 and put M​use Spark Open on Hugging Face three days later, on October 15. They are not two versions of the same product. Muse Wonder is a closed frontier model you reach through an API or a subscription; M​use Spark Open is a gated download that arrives as 816 billion parameters of mixture-of-experts weights, a licence, and a hardware decision. Comparing them on a leaderboard is therefore the wrong first move, because the leaderboard measures capability and the thing that separates them is who owns the machine. This piece works through both — the measured gaps, what they cost per unit of work, and the four situations in which the answer flips — using the numbers that are actually published, with a note wherever a number is a vendor claim rather than a measurement.

相隔三天，而且不屬於同一個類別

The release order matters more than the release dates. Muse Wonder shipped first as two checkpoints: muse-wonder-1.0, served on Meta's Model API and inside Muse Code, and muse-wonder-full-1.0, which Meta sells only through a Muse Code subscription and does not expose as an API model. M​use Spark Open shipped next as a single artefact: a BF16 checkpoint and an NVFP4 build behind an access-controlled repository, a model card, a technical report, and the Muse Model License. One of those is a service. The other is a file.

後續一切都是由這一點推導而來。封閉模型的價格完全由供應商說了算，而其能力也可能隨時被供應商修訂。開放模型的價格則是一個資本數字——GPU、電力、工程師——而它的能力取決於你的推論服務堆疊能從中榨取多少。因此，有用的比較不是「哪個分數更高」——這根本不是接近的競爭——而是「這兩者之中，哪一個能以更低成本讓你達到所需能力」，而這完全取決於你已經擁有什麼，以及你被允許對權重做什麼。

兩個模型都接受文字、圖像、語音和影片輸入，並輸出文字，且兩者都宣稱具備 1,048,576 個 token 的上下文視窗。相似之處僅止於此。

計分板今天的實際狀況，而非你記憶中的樣子

Artificial Analysis 重新調整了其 Intelligence Index，目前版本為 v5.0.1，納入九項評測：AA-Briefcase v1.1、AutomationBench-AA、Terminal-Bench 4.0、FrontierCode-AA 2.0、AA-Omniscience v1.1、GDP.pdf、HLE-Diamond、CritPt v2 的困難子集，以及 Terminal-Bench-Science 0.1。這對本次比較很重要，因為在此修訂前你讀到的關於任一模型的任何數字，都與下方數字不在同一尺度上。兩個項目都在同一套件內重新評分，因此它們之間的差距具有可比性，儘管其絕對值與你可能看過的任一模型較舊數字不可比。

On that revision, M​use Spark Open (Xhigh) scores 25 against a median of 23 among other open-weight models of similar size — above average, but not by much. Muse Wonder (Max) scores 44 against a median of 40 among models in a similar price tier. Nineteen points is not a rounding difference, but it is also not the whole story, because the two medians describe two different comparison sets: one is a small class of same-size open models, the other a broader and more expensive tier. What the pair does establish cleanly is that each model is above its own peer group, which is a more useful statement than a rank and considerably harder to game.

每次評估的資料列以更高的解析度表達相同內容。下列每一行都在同一維度上、以相同順序涵蓋兩側，因此可以成對閱讀，而不是當作兩項各自獨立的陳述。

• Index, v5.0.1 — M​use Spark Open 25 at Xhigh effort, against a class median of 23; Muse Wonder 44 at Max effort, against a class median of 40

• AA-Briefcase v1.1 (agentic knowledge work, Elo) — 1,278.7 for M​use Spark Open; 1,658.7 for Muse Wonder

• AutomationBench-AA (agentic SaaS workflows) — 0.497 for M​use Spark Open; 0.681 for Muse Wonder

• Terminal-Bench 4.0 — 0.081 for M​use Spark Open; 0.520 for Muse Wonder

• FrontierCode-AA 2.0 — 0.076 for M​use Spark Open; 0.240 for Muse Wonder

• HLE-Diamond — 0.263 for M​use Spark Open; 0.390 for Muse Wonder

• CritPt v2, hard subset — 0.295 for M​use Spark Open; 0.505 for Muse Wonder

• GDP.pdf, all-pass — 0.186 for M​use Spark Open; 0.306 for Muse Wonder

• Terminal-Bench-Science 0.1 — 0.019 for M​use Spark Open; 0.257 for Muse Wonder

• AA-Omniscience v1.1 accuracy — 25.8 for M​use Spark Open; 25.1 for Muse Wonder, and the hallucination rate runs the other way: 28.0% for M​use Spark Open against 20.9% for Muse Wonder

Two rows deserve a sentence each. Terminal-Bench-Science at 0.019 against 0.257 is the widest proportional gap on the sheet, and Terminal-Bench 4.0 at 0.081 against 0.520 is the widest on the workload both models are actually marketed for — an agent that runs commands. If that is your use case, this is close to the whole argument. The Omniscience pair is the one place the cheaper model holds its ground: on knowledge accuracy the two are within a point, and the thing that separates them is not how much they know but how readily they invent an answer when they do not — roughly one in 3.6 responses for M​use Spark Open against one in 4.8 for Muse Wonder.

對這一切有個但書。這些是評測方的數字，不是廠商的數字，這使它們成為現有兩類證據中較強的一類。它們仍然只是單一實驗室的測試框架，跑的是本身在本週期才修訂過的套件，而且其中幾項評估夠新，評分還會持續變動。把方向視為可靠，把小數點後第三位視為雜訊。

參數數量，以及紀錄究竟到哪裡就戛然而止

M​use Spark Open publishes its size as part of being open: 816 billion total parameters with roughly 42 billion active per token, 62 decoder layers, 256 experts per mixture-of-experts layer with 8 selected per token, and experts held in 4-bit NVFP4 in the shipped quantised build with some layers left in bf16. The 42-billion-active figure is the one that matters for serving, because it sets how much weight has to be read for every token generated, and it is the reason the model can be served on eight GPUs at all.

就以 Muse Wonder 來說，我們能查核的紀錄上並沒有對應的數字。Artificial Analysis 將其標為專有，而其技術規格面板列出了推理、輸入模態、輸出模態與上下文——完全沒有總參數或有效參數列。我們不會從價格或從基準測試推斷參數數量，也不會重複一個我們無法指出的數字。可以說的是方向性的：一個位於 $4/$20 費率卡上的 44 指數封閉模型，所服務的規模明顯大於一個 420 億有效參數的開放模型，而它承擔的每一項服務成本都是別人的資本，而不是你的。

便宜的模型同時也是速度快的那個，這點令人意外。

價格與速度通常此消彼長。在這裡，兩者卻同向而行，而這是兩頁中任一頁最具決策相關性的發現。

• Output speed — 169 tokens per second for M​use Spark Open; 118 tokens per second for Muse Wonder

• Time to first answer token — 13.81 seconds for M​use Spark Open; 25.19 seconds for Muse Wonder

• Time per Index task — 6.6 minutes for M​use Spark Open; 13.6 minutes for Muse Wonder

• Output tokens used in the Index run — 90.6 million for M​use Spark Open; 113 million for Muse Wonder, against a median of 72.8 million for comparable models

Read the last two lines together and the mechanism is obvious. Muse Wonder is a reasoning model in the strict sense: it works longer, produces more tokens, and takes roughly twice the wall-clock time per task. M​use Spark Open, at Xhigh effort, is leaner — it generated fewer output tokens than the median comparable model, which is unusual for a reasoning configuration and is a genuine mark against the assumption that more thinking is always more capability. The 169 against 118 figure is a hosted-endpoint measurement including time to first token, so it is a statement about the deployed endpoints rather than about the silicon underneath either model.

實際後果是，價格差距比費率表所顯示的更大，因為你按 token 付費，而 {{1}}Muse Wonder{{/1}} 每項任務會輸出更多 token。這就帶出唯一能在不同價格與不同冗長程度下，誠實比較兩個模型的數字：每項任務成本。

每個任務實際上要花多少成本

Artificial Analysis 會計算執行其整套評估所需成本，再除以任務數量，並依每項評估對指數的貢獻進行加權。如此得出的同基準數字，已經吸收了冗長度差異。

• Cost per Index task — $2.02 for M​use Spark Open; $6.95 for Muse Wonder

• Blended list price, 3:1 input to output — $2.00 per million tokens for M​use Spark Open; $8.00 per million for Muse Wonder

• Cache read — $0.15 per million tokens for M​use Spark Open, an 88% discount; $0.20 per million for Muse Wonder, a 95% discount

The headline rate card is $1.25 in and $4.25 out against $4.00 in and $20.00 out, so the output price differs by 4.7× and the input price by 3.2×. The realised per-task cost differs by 3.4× rather than 4.7×, which is the verbosity and cache effects doing their work: Muse Wonder's 95% cache discount is steeper than M​use Spark Open's 88%, and heavy caching pulls its effective cost back toward the cheaper model's. If your workload is cache-friendly — long shared system prompts, repeated documents — the rate card overstates the gap. If it is not, the per-task number is the one to plan against and the rate card understates nothing.

Divide each per-task cost by its Index score and you get a third number, which is mine rather than the evaluator's: about eight cents per index point for M​use Spark Open against about sixteen cents for Muse Wonder. That is the cleanest statement of the trade. You are paying roughly twice as much per unit of measured capability for the closed model, and what you receive for the premium is nineteen index points and a lower hallucination rate.

授權費是價目表隱藏的一項成本。

M​use Spark Open's terms are the Muse Model License, and two of its clauses change what "free weights" means in a business plan.

第一項是模型即服務條款。如果你透過 API 將模型作為產品或服務提供給第三方存取，且你在任何連續十二個月期間內從中取得的合計總收入超過兩千萬美元，你必須通知 Meta，並在繼續之前簽訂另行的書面協議。將模型嵌入自家產品屬於例外；轉售則不屬之。打算以這些權重為基礎成為推論服務供應商的團隊，應該在撰寫架構文件之前閱讀該條款，而不是之後才讀。

The second is attribution: the licence requires "Built with M​use Spark Open" to be displayed prominently in the interface of any product using the model or a derivative, or in its documentation, site or settings. That is a small tax and an unconditional one, and it appears exactly where a product decision gets made — the about screen.

授權條款明確賦予的，正是讓下載站得住腳的那一部分：微調、蒸餾、合併、發布、再授權與販售衍生作品，且 Meta 聲稱不擁有輸出成果的所有權。Muse Wonder 這些一概不提供。你無法微調託管檢查點，無法把它合併進領域模型，無法在隔離環境中運行它，也無法透過重新訓練來修正某個行為。你為 Muse Wonder 支付的溢價，買到的是能力，也讓你免於營運負擔；但它同時也讓你無緣所有需要動到權重的選項。

有一個容易讓人卡住的實際問題：下載本身是有門檻的。這個儲存庫設有存取管制，申請表要求填寫法定姓名、出生日期，以及附有公司識別資料的機構名稱，而且送出後無法修改。請預留排隊等待的時間，而不是以為按一下就好。

在 megakernel 改變算術之處

There is a third number in this comparison that neither rate card contains, and it arrived yesterday from outside Meta. On November 11, 2026 the SGLang team published a megakernel for M​use Spark Open — one resident kernel per layer instead of roughly 1,450 launches per decode step — and measured decoding going from 517 to 1,109 tokens per second on eight GB300s. That is a 2.1× improvement on the same weights, the same precision, and the same numerics, achieved purely in software.

把這一點拿來和代管服務的數據並列，這項決策的形貌就會改變。租用的端點每秒回應 169 個 token；一套自行打造、執行研究用核心的八 GPU 堆疊則每秒回應 1,109 個。第二個數字是在 batch size 為 1、單一 GSM8K 請求、取 20 個步驟的中位數下測得，而且除了作者之外，沒有人重現過這個結果——所以那是你可能達到的上限，而不是能放進容量規劃的吞吐量數字。但它建立了定義這組模型的非對稱性：封閉模型的服務效率由 Meta 決定要不要改善，而你只能等；相對地，開放模型的服務效率則是一個你可以出動自家工程師加以攻克的函數。

對於 GPU 已經付清的團隊來說，在開源模型表現接近的工作負載中，這種不對稱性的價值超過十九個指數點。對於沒有這些 GPU 的團隊來說，這毫無價值，因為 1,109 這個數字是附在一個你必須購買的八 GPU 節點上。

四種情況，四種不同答案

只要點明限制條件，比較就能清楚地得出結論，而這個限制幾乎從來不是基準。

如果你這一季要推出產品，而工作內容是高難度推理、長時間跨度的代理式任務，或任何你的問題形態就是終端機基準測試中 0.081 對 0.520 的差距，那麼答案就是 Muse Wonder，而且毫無懸念。你每項任務支付 6.95 美元，換來的是能完成任務的模型；另一種選擇則是每項任務支付 2.02 美元，換來一個經常無法完成任務的模型。

If you are running high-volume, latency-sensitive work where an answer at 169 tokens per second with a first token in 13.8 seconds is acceptable and one at 118 tokens per second with a first token in 25.2 seconds is not, the answer is M​use Spark Open even before you consider price — and the price then makes it a 3.4×-cheaper answer. The 28% hallucination rate is the thing to design around: ground it, cite it, or verify it downstream.

If you have data that cannot leave your infrastructure, or a regulator who wants to see the weights, or a domain where fine-tuning on your own corpus is the entire value proposition, the comparison is over. Muse Wonder cannot be that, and no amount of index points changes it. M​use Spark Open can, and the licence explicitly permits the fine-tune, the merge and the redistribution.

如果你已經擁有配備 GB300 等級節點的訓練叢集，開源模型的經濟效益根本不是價目表上的費率，而是你已經攤銷完的 GPU 的邊際成本。這正是每秒 1,109 個 token 舉足輕重的情境，也是 megakernel 相關工作可直接付諸實踐的情境，更是那份兩千萬美元的模型即服務條款成為阻擋在副業專案與正規事業之間的唯一一道文書關卡的情境。

無法解決的那種情況，正是人們最常問到的：一個小團隊既想要前沿能力，又想要便宜。它無法從這對模型中同時得到兩者。它能得到的是串接式做法——用便宜模型處理大部分流量，用昂貴模型處理剩下的部分——而這是路由決策，不是模型決策。

我們會路由什麼，以及我們不會路由什麼

Neither of these two models is on OrcaRouter. There is no meta/muse-wonder entry and no meta/muse-spark-open entry in the catalogue, and we are not going to describe either as available through us. M​use Spark Open is a download and a licence; Muse Wonder is Meta's own endpoint and a subscription.

What we do carry from the same lab is the M​use Spark line: Muse Spark 1.2 and Muse Spark 1.1. That matters for exactly one of the four situations above — the cascade. A routing layer is where the "cheap model for most traffic, expensive model for the residue" answer stops being an architecture diagram and becomes configuration: one API across 200+ models, with the provider's list price passed through at 0% markup, so a Meta price cut lands on your bill the same day rather than at the next contract renewal, and automatic failover plus a routing DSL that moves traffic between models by configuration rather than by a code change and a deploy.

就開源權重那一半的比較而言，老實說，沒有任何路由器幫得上忙。權重是你的、GPU 是你的、核心也是你的，而要達到 1,109 這個數字，唯一的途徑就是一位帶著效能分析工具的工程師。

當這項比較變得過時

• Muse Wonder Full 提供 API 存取。得分較高的檢查點目前被封鎖在訂閱服務內，因此本次比較中的 Muse Wonder 是得分 44 的 API 模型，而非 Meta 另行公布的 Full 檢查點。若 Full 開放端點，能力差距將會擴大，而串聯論點也會更難成立。

• 獨立重現 megakernel 的數據。某個團隊在批次大小為一的情況下達到 2.1×，這只是上限；第二個團隊在真實批次大小下也能達到相同水準，就讓它成為附有數據的自我託管案例。

• A price move on either side. M​use Spark Open's hosted rate is $1.25/$4.25 and Muse Wonder's is $4.00/$20.00, and a single vendor announcement could compress or invert the 3.4× per-task gap. This is the cheapest thing on the list to change and the most likely to.

• A fine-tune that closes the gap. Nineteen index points is a large deficit for a derivative to recover, and the open model's licence is what makes the attempt legal. If a domain tune on M​use Spark Open lands near the median of Muse Wonder's class on the specific evaluations your product cares about, the premium stops being justified for that product.

Until one of those moves, the summary is short. Muse Wonder is the more capable model and you rent it by the token. M​use Spark Open is the less capable model and you own it outright, along with the GPUs, the licence clauses, the hallucination rate, and — as of this month — a public recipe for making it twice as fast.