Moondream 3.1 9B A2B 對比 UI-Venus 2 9B：一方已公佈價格，一方授權待定

M​oondream 3.1 9B A2B and UI-Venus 2 9B are the same size, in the same modality, and they hand back the same primitive — a coordinate. Only one of them has a published per-token price, and the price belongs to the one whose hosted endpoint is about to be switched off. M​oondream 3.1 9B A2B is sold at $0.30 per million input tokens and $1.00 per million output tokens on M​oondream Cloud. On 13 November 2026, four weeks from today, M​oondream Cloud — inference and the Lens fine-tuning product together — shuts down, a line that arrived on 14 October alongside the news that M​oondream the company has been acquired by Prime Intellect. UI-Venus 2 9B has no first-party endpoint at all, so it has no price list either. And the licence that would tell you whether you may ship it commercially is still, in the words of its own model card, pending final confirmation.

Neither of these models is new, and nothing here is a launch. M​oondream 3.1 9B A2B shipped on 7 July 2026 under the M​oondream Model License 1.0. UI-Venus 2 9B went up on Hugging Face on 26 August 2026 from inclusionAI, the Venus Team's repository at Ant Group, with its technical report following on 27 August. What actually moved inside the last seven days is on the M​oondream side: the acquisition, dated 14 October on M​oondream's own blog and a day earlier on Prime Intellect's, and the thirty-day wind-down of the hosted service that came attached to it. That is the reason this pairing is worth reading now rather than in a launch round-up. One of these two models just had its distribution channel put on a countdown; the other never had one.

你今天可以撥打什麼，以及可以打到什麼時候

M​oondream 3.1 9B A2B is available three ways, and one of them is on a clock.

廠商的雲端服務按 token 計費：即時延遲下每百萬個輸入 token 為 0.30 美元、每百萬個輸出 token 為 1.00 美元；若使用批次 API，則分別為 0.15 美元與 0.50 美元，在牌價之上再享 50% 折扣。微調產品 Lens 的收費為每百萬個訓練 token 0.60 美元。存取權限是受到節流限制，而非依併發數計量——廠商的速率限制頁面載明，任何 API 金鑰每秒可發 2 次請求，一旦帳戶內有 10 美元，便提升至每秒 10 次請求。接著是期限：2026 年 11 月 13 日，雲端推論端點與 Lens 都會關閉，未使用的已購買點數將獲退款，客戶也會被告知將協助他們遷移。

The weights, meanwhile, do not go anywhere. M​oondream Model License 1.0 permits commercial self-hosting, fine-tuning, quantisation and redistribution, and Photon — the vendor's own inference engine — runs the checkpoint locally on N​VIDIA Ampere-or-newer hardware and Apple Silicon at no charge. So the deadline splits cleanly. If your code posts to the vendor's hosted endpoint, you have a migration in front of you and roughly four weeks to do it. If your code loads a checkpoint you own, the announcement changed nothing you have to act on.

UI-Venus 2 9B 從來無法以那種方式呼叫。它是從 Qwen3.5-9B 初始化而來的稠密 9B 模型的全參數權重，拆分為四個 BF16 safetensors 分片，而預定的使用途徑是你自行架設的 vLLM 伺服器，並以 262,144 個權杖的最大模型長度及 Qwen3 推理剖析器啟動。這張模型卡對這套作法的限制異常直白：服務配置「並未作為本次模型卡更新的一部分進行線上金絲雀驗證，因此部署前請固定並驗證適合 Qwen3.5 的 vLLM 版本。」

把這兩段並排放在一起，買家看到的差異就成了一筆明細支出，而不是一項基準測試。其中一個模型是按 token 計費的帳單，你可以放進試算表一路算到十一月中。另一個則是一顆你得自己找到、決定規格、還得讓它保持熱機的 GPU——而發布權重的廠商從未替你驗證過伺服器配置。

參數數量相同，token 帳單卻截然不同

9B 標籤是這兩者共有之中最沒用的東西。真正在經濟效益上區分它們的，是每項任務的回合數。

M​oondream 3.1 9B A2B has no second step. One call is one image in and a short answer out — bounding boxes from detect, centre coordinates from point, a caption, or a written answer from query. There is an optional reasoning mode, and the vendor's documentation is blunt about its cost: enabling it generates an explicit analysis before the answer and can "materially increase latency and generated tokens." The output is deliberately small, and the model's whole pitch is that it is small while remaining useful.

UI-Venus 2 9B 是一個閉環。每一步會取得當前畫面，帶著該回合的截圖歷史，輸出推理軌跡，接著輸出一個動作，而下一個截圖則會回饋成為輸入。已發表的評估顯示這在回合數上的代價：模型卡報告了 MobileWorld 在 50 步預算下的結果，又在 100 步下報告一次，而 OSWorld 2.0 則是在 150 步預算下。在這套設計中，上下文並不是容納一份文件的空間——而是這個代理還能記得多少個螢幕狀態。

The arithmetic follows, and it is ours rather than the vendors'. If a caption or a detection answer is 50 to 100 output tokens, then a million output tokens is somewhere between ten and twenty thousand M​oondream calls for a dollar. You cannot write the equivalent sentence for UI-Venus 2 9B, because there is no price to divide by — you pay for the GPU hours the loop occupies, not for the tokens it produces, and the token count per completed task is a function of how many steps you allow. Same parameter count, same 9B headline, and the two bills are not in the same unit.

One more asymmetry worth knowing before it bites a production plan. M​oondream measures its own engine at 34.2 requests per second on a single H100 SXM5 at batch size 16 — a vendor figure, on the vendor's hardware, with Photon serving. The hosted API caps a funded account at 10 requests per second. That is a three-fold gap between what the model can do on hardware you control and what the endpoint will let you do, and it is the sort of thing a pipeline discovers in production rather than in an evaluation.

並排對比，逐項比較

• Task — M​oondream 3.1 9B A2B: perception over a single image. UI-Venus 2 9B: agency over a screen, in a loop.

• 參數 — 總計 9B，每個 token 有 2B 處於活躍狀態，為稀疏混合專家架構。UI-Venus 2 9B：9B 稠密，從 Qwen3.5-9B 初始化。

• 輸出 — 邊界框、中心座標、圖說、書面答案與分割路徑。UI-Venus 2 9B：一個動作、一個點擊座標，以及產生它們的推理軌跡。

• 上下文 — 32K 個符元。UI-Venus 2 9B：262,144 個最大位置嵌入，設定為情節記憶，而非文件上下文。

• 公佈價格 — 即時每百萬個符元輸入 $0.30、輸出 $1.00，批次為 $0.15 與 $0.50，適用於將於 2026 年 11 月 13 日關閉的端點。UI-Venus 2 9B：無，因為沒有第一方端點。你要自行託管，否則就不要使用。

• 已發布的基準測試 — COCO 81.46 F1@0.5、ODinW-13 93.94、RefCOCO-M 87.1、CountBench 90.35、ChartQA 86.01、DocVQA 88.6，以及套件平均 77.9。UI-Venus 2 9B：AndroidWorld 80.2、WebVoyager 90.8、OSWorld-Verified 70.8、ScreenSpot-Pro 73.0、VenusBench-CAPTCHA 78.1，以及 OSHarm 攻擊成功率 11.3%，相較於其基礎模型的 25.3%。上述每一項數字皆出自廠商本身。

• Licence — M​oondream Model License 1.0: open weights, commercial self-hosting permitted, redistribution permitted, one restriction. UI-Venus 2 9B: pending final confirmation, with no licence file in the repository.

• 獨立評分 —— 無。兩個模型都沒有登上中立排行榜的紀錄，而兩家廠商也沒有發表任何共同的基準測試，因此上表兩欄中的任何內容都無法相互對照。

許可證正是兩者不再相似的地方。

M​oondream's licence is short and specific, and the vendor rewrote it for this release because the previous one was hard to read. Commercial use, self-hosting, fine-tuning, quantisation and redistribution are all permitted without strings. The single restriction is that you may not turn around and offer general-purpose M​oondream inference or fine-tuning to others as a hosted service; a narrow, domain-specific application built on the model is explicitly fine. For a team shipping a product that happens to see images, that is a permissive licence with a competitor clause, and the clause does not touch them.

UI-Venus 2 9B 的立場則不是同一類陳述，值得直接引述，而非改寫轉述。該卡片的 Release Status 區塊指出，模型權重授權「尚待最終確認，並將於公開發布前補上」，並解釋為何未採取那個顯而易見的選項：「Apache-2.0 的聲明刻意未沿用，因為目前上游資料中含有互相衝突的授權聲明。」該儲存庫中沒有任何授權條款檔案。權重於 2026 年 8 月 26 日發布，而該儲存庫最後一次異動是在 9 月 2 日。

實際的解讀並不是這個模型不能用。而是：未釐清的授權是一項商業決策，必須有人白紙黑字地做出決定，而不是你下載一個檢查點就自動繼承的預設狀態。研究、評估與內部原型開發並不是風險所在。把它放在付費客戶面前才是，而上一節的任何基準測試分數都改變不了這一點。這張模型卡還附帶第二項條件，值得任何人在據以開發之前詳讀：CAPTCHA 破解能力「應僅在經授權的環境中使用」——這是對使用情境的約束，而不是關於它的建議。

兩家供應商都讓你從兩種不同方向核對號碼

大多數模型發表只會公布一張表，然後就點到為止。這兩者都公布了但書，而這些但書是不同性質的東西。

M​oondream's release post carries a methodology footer that pins the hardware and the harness: a single H100 SXM5, batch size 16, the highest accuracy and reasoning settings for each model, Photon for M​oondream itself, vLLM for the Qwen3.5 and Gemma 4 baselines, Hugging Face Transformers for SAM 3, and LocateAnything evaluated in slow mode to maximise its accuracy. It also volunteers something a vendor has no obligation to volunteer: the scores were reached by treating each benchmark as a customer task, training a per-task LoRA through the same fine-tuning product customers use, and then folding those LoRAs back into the base weights with on-policy distillation. Read that carefully and it changes what the table describes — the checkpoint you download is a distilled student of its own task specialists, not a model that happened to generalise that well on a task nobody trained it for.

UI-Venus 2 技術報告，arXiv 2609.00028，於 2026 年 8 月 27 日提交，做的是另一種揭露。標有星號的基線是由作者依其自行陳述的協定所評估或重現，而非引用自其他地方。破折號代表未報告，而作者把破折號留著。MobileWorld 印了兩次，一次是 50 步的預算，一次是 100 步，好讓你看出多出來的步數換到了什麼。OSWorld-Verified 附有一項警告，指比較「可能使用各模型專屬的動作鷹架，因此應將其詮釋為基準層級的參考，而非受控的消融實驗」。VenusBench-CAPTCHA 是對 219 個固定範例取的微平均；MCA-Bench 的數字則採用 1,000 個抽樣範例。文中有一句話指出，實境環境的結果可能隨評估日期與環境狀態而異。

這一切都不能讓任一方提出的數字在你的資料上可重現，而兩家供應商也都沒有獨立的結果可以拿出來佐證。它真正代表的是：雙方都告訴了你他們的數字值多少——這比一個漂亮的分數更罕見，對買家也更有用。

以限制條件來看，是哪一個？

• You need boxes, centre points or counts over images, at a price you can forecast — M​oondream 3.1 9B A2B, with the caveat that the forecast only holds if you self-host after 13 November or move the endpoint. A per-token bill with a shutdown date is a forecast with an expiry, not a cost model.

• 你需要一個能讀取螢幕並據此操作的東西 —— UI-Venus 2 9B，而授權問題會比基準測試表更早浮上檯面。先把它用於內部自動化，等授權確定後再重新評估，而不是在採購期間才發現問題。

• You are deploying to hardware you already own — M​oondream 3.1 9B A2B. Photon is free, runs on Ampere-or-newer N​VIDIA parts and Apple Silicon, and gives you the same model in development and production instead of two different endpoints that behave differently under load.

• You want the longer context on the spec sheet — UI-Venus 2 9B's 262,144 tokens is real in configuration, but the intended consumer is a screenshot history, not a PDF. M​oondream's 32K is a quarter of the size and its vendor has never claimed it needs to hold an agent's episode, so neither number is a straight win.

• CAPTCHA 處理已列在你的清單上——請先閱讀卡片的限制。能力已經具備，而授權正是擋在它前面的那個東西。

如果誠實的答案是兩者皆是，只是負責不同的工作——感知在擷取路徑上，而代理在行動路徑上——那就是常見的形態，而且它先是兩個契約的問題，之後才是模型的問題。

模型背後的模型，以及是誰在路由它

這兩個檢查點都不是透過 OrcaRouter 提供服務的，所以這裡並不是在推銷要透過我們來呼叫它們。真正需要路由的，是位於它們後面的那一層。

偵測結果必須經過分流，代理程式提出的動作必須受到評估，而紀錄也必須被寫入。這些都是文字工作，以每個 token 計算，它們是較昂貴的那一半，而這些正是單一 OpenAI 相容金鑰橫跨 200 多個模型——供應商牌價以 0% 加成直接轉嫁，並可跨供應商自動容錯移轉——能從管線中移除一份合約與一條程式碼路徑的所在。當你已經知道想要的推理形態時，路由 DSL 會將數個模型組合成一次呼叫，而模型融合會讓一組模型在同一份輸入上一起運行。兩者都是在不更動與相機通訊的程式碼之下，變換固定視覺模型背後那一層的方法。

在這個特定的比較中，定價機制比平時更為重要。因為標價是直接傳遞，而非被加價，供應商的價格變動當天就會在我們這邊生效。對於一個其託管價格已寫入關閉日期的模型而言，那是唯一值得用來建立預測的定價安排。

什麼會改變這個答案

三件事，而且這三件全都可以查證。

The first is UI-Venus 2 9B's licence resolving. If it lands permissive, the calculus for anyone who wanted the agent but could not justify the legal exposure shifts immediately, and the recommendation above about keeping it internal becomes stale the week it happens. The second is a first-party endpoint for M​oondream 3.1 9B A2B reappearing after the cloud shutdown, which would restore the price comparison this article could only half make. The third, and the one worth the most, is somebody running both models on the same images. Neither vendor has an independent result and the two benchmark sets do not intersect, so the first person to score these two on their own data produces the only comparison that will settle the grounding question — and it is an afternoon of work, not a research project.

在那之前，站得住腳的摘要刻意保持狹隘。這些是兩個 9B 視覺模型，共享相同的參數量、模態標籤，以及回傳座標的習性，而它們對於決定部署幾乎沒有共識：一次呼叫的成本是多少、一次呼叫是否甚至是正確的單位，以及你是否被允許把輸出交付出去。

這些都不是假設：像這次發布這類版本背後那些有紀錄的 AI 失效與代理事件，全都收錄在OrcaRouter AI Incident Archive，其中每筆條目都附有自己的來源、嚴重程度、信心等級與爭議標記。