一張生成的標題卡寫著 Qwen4-Max - 合作夥伴卡,逐行閱讀,並將三個數字排成費率列:每百萬個 token 輸入 $1.413、輸出 $4.24、快取讀取 $0.071,標示為全球範圍費率,旁邊還有一個標籤寫著 262,144 共用預算。
Guides & Insights

Qwen4-Max 的新合作夥伴卡宣傳四項能力。其中三項日後會讓你付出代價。

作者

jinhao.song

發佈日期

最新模型 · 20查看全部模型 →
基準測試:Artificial Analysis · 每日更新
返回全部文章

A second platform is now selling Qwen4-Max, and its announcement reads like a product launch: a 1M-token context, text, image and video input, tool calling, structured output, reasoning, and $1.413 per million input tokens, $4.24 per million output, $0.071 per million cached. The model went generally available on A​libaba's own API on 16 November 2026, so a second seller two days later is fast but not surprising. What is worth reading closely is the feature list, because three of those four capabilities behave differently on a partner platform than they do on A​libaba's own, and the price is the clue that tells you why.

Start with the price, because everything else follows from it. $1.413 and $4.24 are not new numbers. They are A​libaba's Global-scope rates — Beijing, Hong Kong, Frankfurt, US Virginia and Tokyo — to three decimal places, sitting 5.8% below the Singapore card of $1.50 and $4.50. The cached rate is the tell: $0.071 is the Global figure, and Singapore's is $0.075. Nobody arrives at $0.071 independently. A partner card quoting the vendor's regional rate to the thousandth of a dollar is quoting a pass-through, and a pass-through means one deployment of the model with a second billing layer on top, not a second training run or a second set of kernels.

那個區別並非咬文嚼字。它為卡片上其餘每一列設定了預期:價格在直接傳遞後依然成立,行為則不然——除非合作夥伴已投入工程開發,讓它得以成立。

A generated two-column comparison card titled Qwen4-Max - the advertised card, checked, with the partner-listing column reading 1M tokens context, output ceiling not stated, $1.413 input, $4.24 output, video input yes and reasoning dial not stated, against the vendor column reading 1M with 984k usable, 262,144 shared, $1.413 Global, $4.24 Global, video yes, and low, high and xhigh on the dial.

主張一:1M 上下文。沒錯,但仍是這頁上最沒用的數字

Every Qwen4-Max listing advertises a million tokens, and every one of them is technically right. A​libaba's own documentation gives 1,000,000 tokens of context with a 991,808-token input ceiling when thinking is off and 983,616 when it is on, and Artificial Analysis independently records a 984k-token context window. Those three numbers are the same 1M with different amounts of the truth exposed.

真正關鍵的那一列不是輸入上限,而是輸出預算,而這個類別中沒有任何合作夥伴卡會標示這項資訊。Qwen4-Max 有 262,144 個 token 的預算,由思考與最終答案共用;其中關閉思考時可用 131,072 個。在 900,000 個 token 的輸入下,這表示在思考模式下,視窗關閉前你最多只剩 91,808 個 token 可用於回應——而且推理 token 也計入其中。團隊若以「1M 輸入、若干輸出」來估算長上下文工作,就是在拿廠商從未公布的數字來估算。

這是合作夥伴上架資訊可能悄悄讓你花錢的第一個地方。一張標示 1M 的卡片,完全沒說明平台是否會強制執行共用預算、是否會截斷推理內容而不是答案,或是否會在輸出達到 128K 時默默停止。這些失敗模式會以凌晨兩點出現被截斷的 JSON 物件的形式顯現,而不是以錯誤碼的形式。

主張二:視訊輸入。這一列決定這個商品刊登是不是真正的連接埠。

文字、圖像與影片輸入是本次公告中最難以造假的能力。要向 Qwen4-Max 提供影片服務,並不是接受一個 URL 再把它轉送過去那麼簡單:必須對影格進行取樣,並以模型訓練時所用的解析度與影格率,將其編碼進模型的視覺 token 串流,還得計入一項本來就已屬稀缺資源的上下文。只接通了文字路徑的平台,要嘛直接拒絕影片,要嘛接受影片卻悄悄降低其品質。

Which is why the video row is the one to test first if you are evaluating a partner. A one-minute 30fps clip at a conservative sampling rate is thousands of visual tokens; drop it into a 700,000-token document prompt and you are at the input ceiling with a shorter answer than you planned for. Qwen4-Max's video support is real on A​libaba's own API — the model is documented for it, and Artificial Analysis records text, image and video as its modalities — but a partner card is an assertion, not a test result.

圖像路徑也有同樣問題,只是程度較輕。解析度處理、圖磚切割,以及高 DPI 螢幕截圖的 token 成本,全都是平台層面的決定,而這些決定沒有一項會出現在以每百萬個 token 為輸入計價的價目表上。

主張三:工具呼叫與結構化輸出。這是合作夥伴唯一能勝過供應商的地方

Structured output and tool calling are where a pass-through platform can add real value, and it is worth saying so plainly because the rest of this piece is cautious. A​libaba documents structured output, parallel tool calling, prefix completion, web search and both explicit and implicit caching across four protocol surfaces — O​penAI Chat Completions, O​penAI Responses, A​nthropic Messages and DashScope. A partner that exposes only the O​penAI shape is offering less surface, not a worse model.

不過,若合作夥伴在其上建構了受結構描述約束的解碼器,就能提供供應商自家端點無法保證的東西:每次都能在第一次嘗試就解析成功的 JSON,無需重試迴圈。對於大量從發票擷取欄位的管線而言,每避免一次重試,就等於節省 2 倍的輸入 token,而這遠遠超過新加坡與 Global 價格卡之間 5.8% 的區域差異。

所以關於這項能力,該問的問題不是「是否支援結構化輸出」——每一份清單都說有——而是「它是受限解碼,還是提示指令」。前者是保證;後者只是強烈建議,每五十次會失敗一次,而且通常發生在整批請求中最長、最昂貴的那一筆。

主張四:推理。最難移植的能力,也是沒有人定價的能力。

Qwen4-Max 出貨時即開啟推理,並將 effort 預設為 high。此調節項可接受 low、high 與 xhigh;思考預算預設為 131,072 個 token,而在思考模式下,max_completion_tokens 預設為 262,144。若請求未設定這三項中的任何一項,就會預設採用昂貴的設定,而帳單大部分正是來自這個預設設定:Artificial Analysis 測得每個索引任務有 104,940 個輸出 token,其中 68,914 個是推理;相較之下,整個九項評測套件的領域中位數為 6,510 萬個輸出 token。Qwen4-Max 則產生了 1.27 億個。

On a partner platform, three things can happen to that dial and only one of them is visible. The reasoning effort parameter can be passed through unchanged, in which case everything above holds. It can be accepted and ignored, in which case your low effort request bills like a high one. Or it can be rejected, in which case you find out with a 400 and adjust. The first is fine, the third is annoying, and the second is the one that quietly doubles a monthly invoice — and it is the most common, because passing an unknown field through to a vendor API is often the default behaviour of an O​penAI-compatible proxy.

這也是價格宣稱受到最嚴格檢驗的地方。在 $1.413 / $4.24 的情況下,合作夥伴是按供應商的 Global 費率計費,因此 token 價格是誠實的。token 數量是否誠實,則取決於一個未顯示在卡片上的參數。

一則「合作」公告能告訴你什麼,以及它不能告訴你什麼

The framing of this launch — a partnership with A​libaba's Qwen team rather than a resale — is worth taking at face value for what it covers and not a word more. It plausibly covers early access to weights, capacity allocation and a supported identifier. It does not cover the things that decide your bill: the thinking budget, the input ceiling in each mode, the video sampling policy, the constrained-decoding guarantee, or which regions the deployment sits in.

There is a testable version of this. Ask the platform for its input ceiling in thinking mode. A​libaba publishes 983,616. If the answer is "1M", the platform has not read the vendor's documentation, and the rest of its spec sheet deserves the same scepticism. That single question separates a partner that has ported the model from one that has ported the price.

獨立資料現在比推出時更好,這有幫助。Artificial Analysis 在 Intelligence Index v5.0.1 上給 Qwen4-Max 評分 43——這是同類模型的中位數——每完成一項索引任務為 $3.27、每項任務 32.1 分鐘、每秒 54.5 個輸出 token,以及首個回答 token 耗時 155.4 秒。那是對一個推出僅兩天的模型所做的真正獨立評估,也是面對任何廠商圖表時該引用的數字。它沒告訴你的是,合作夥伴平台是否會重現這個結果:追蹤器只在一個端點上測量該模型,而第二個部署若採用不同的批次處理、不同的核心或不同的推理預設,結果可能相差好幾分,成本也可能相差好幾倍。

A screenshot of the Artificial Analysis page for Qwen4 Max captured 18 November 2026 with an English interface, showing the Alibaba provider tag and a November 2026 release, an Intelligence Index of 43, output speed of 55 tokens per second, 185.4 seconds to first token, a 984k-token context window, and a summary line describing the model as below average in intelligence, well priced, notably slow per task and very verbose.

如何不靠猜測就買到這個

OrcaRouter does not host Qwen4-Max. The model is reachable through A​libaba's own Model Studio API and several third-party platforms, and that is the honest description of the situation — no vendor-hosted route on this platform, so no claim to one. What OrcaRouter does provide is the comparison point: 200-plus model identifiers behind one API at 0% markup, provider list price passed through, so the $1.50 Singapore card and the $1.413 Global card are the same numbers here as on A​libaba's own pricing page, and a vendor price change is live the same day rather than at the next billing cycle.

A screenshot of OrcaRouter's models catalogue captured 18 November 2026 with an English interface, showing the header '226 models - 17 providers - one API key, one bill', the filter rail for input modalities, context length, input price, status, access tier, series, supported parameters, privacy and inference region, and the first card visible being kimi/kimi-k3.1-pro next to a two-step 'how to call any model' panel.

對正在決定合作夥伴上架項目是否值得採用的團隊來說,實用的做法是並排對照,而不是只看評測。把同樣 50 個請求——一份附有影片的長文件、一批結構化擷取、一條代理式工具呼叫鏈、一個低工作量的分類任務——送到供應商自己的端點,也送到合作夥伴的端點,並比較計費的 token 數,而不是回傳的 token 數。這兩個數字恰恰會在某個參數被丟掉時出現落差,而路由 DSL 能讓你針對合作夥伴確實較佳的工作負載保留它,並在它並非較佳的工作負載上自動回退。

The announcement is accurate. The price on it is A​libaba's own Global rate, the modality list matches the vendor's documentation, and a partner that quotes a cache read to three decimal places has clearly done the arithmetic. What it leaves out is the part you cannot verify from a card: a shared 262,144-token budget that shrinks the usable window, a reasoning dial that decides most of the invoice, a video path that is either real engineering or an accepted field, and a structured-output promise that is either constrained decoding or a wish. Those four things are worth more of your evaluation time than the 5.8% the partner price saves.