Liquid AI Releases d1-ultra: A Multimodal Decision Model That Scores 65 on the AA Decision Index
https://www.marktechpost.com/2026/10/15/liquid-ai-d1-ultra-decision-model/📌 Liquid AI d1-ultra:不輸出任何文字的決策模型,拿下 AA Decision Index 第一
TL;DR:d1-ultra 用單次前向傳播直接輸出答案機率、零輸出 token,專為路由、審核、分類場景設計,API 已開放試用。
如果你呼叫 LLM 只是要一個「是/否」或分類標籤,為什麼要付生成整句話的費用?Liquid AI 的答案是:把輸出壓縮成一組機率分布,連一個 token 都不用生成。
🤔 決策模型:用機率取代自由文字
Liquid AI 釋出了 d1-ultra,是其目前最強的決策模型。它讀取文字、影像或兩者,針對有範圍限制的問題,直接回傳每個可能答案的機率,僅需一次前向傳播,不產生任何輸出 token。Liquid 的決策模型文件描述了 3 種問題類型,並建議將決策模型用於路由、內容審核、守門機制、分流(triage)以及取代 LLM-as-judge 的場景;至於摘要、聊天與程式碼生成,文件仍建議使用一般語言模型。
🧩 一次呼叫評估多個問題
每個請求帶著一個 state(也就是上下文)、可選的影像,以及一個或多個「有型別」的問題,模型會在同一次呼叫中,針對同樣的輸入評估每一個問題。d1 的 API 參考文件顯示,回應中的 output_tokens 永遠為 0。
Liquid AI 團隊表示,d1-ultra 使用比 d1 更大的開源權重骨幹,並疊加了大量 Liquid 專屬的客製化,搭配 Liquid 的決策訓練 pipeline,在知識型任務上的提升最為明顯。團隊舉的應用例子包括:對照規格檢查電路板是否有瑕疵,或是比對客戶照片來核實理賠申請,這類「雜亂」的實務場景正是 d1-ultra 的設計目標。
📊 在 Artificial Analysis Decision Index 排名第一
Liquid AI 將 d1-ultra 與前代 d1,以及 TypeSafe 的 Jev 1.14 在 112 項評測上做比較,根據官方公告,d1-ultra 在相同價格下於所有任務上都勝過 Jev 1.14。在獨立的 Artificial Analysis Decision Index 上,Liquid 回報 d1-ultra 以 65 分排名第一,該指數由 4 項評測各佔 25% 權重平均而成。作為對照,Artificial Analysis 在 10 月 13 日的發布文章中,Jev 1.13 以 58 分排名首位,GLiDE 與 Microsoft-Decision-1 則並列 54 分。
💡 計費只算輸入 token,但影像會被重複計費
d1-ultra 透過 Liquid API 的 /decisions/v1/systemone 端點提供服務,model 參數設為 “d1-ultra”,目前已可在 Liquid API 上使用,定價為每百萬輸入 token 0.042 美元;Liquid 表示 Vercel AI Gateway 與 OpenRouter 的支援即將推出,也可以先在 d1 Playground 試用。
計費只計算輸入 token。根據 Liquid 的定價說明,影像以每個 32×32 像素 patch 計 1.5 個 token,一張 1024×1024 的影像因此要價 1,536 個 token。需要注意的是,每個問題都會被當作獨立的 prompt 計費,包含其文字與所有附帶影像,所以如果對同一張影像提出 3 個問題,這張影像會被計費 3 次。端點文件也設定上限:單次請求最多 8 張影像、10,000 個影像 patch,完整 JSON body 需低於 4.5MB。
🎯 實務啟示
對於分類、路由、內容審核、分流這類「只要一個答案機率、不需要生成文字」的場景,d1-ultra 提供了比一般 LLM 呼叫更直接的成本結構。但在設計請求時要留意:把多個問題疊加在同一張影像上,會讓該影像的 token 成本隨問題數線性增加,批次設計時值得先估算好影像與問題的組合方式,避免費用超出預期。
🔗 來源
- 標題:Liquid AI Releases d1-ultra: A Multimodal Decision Model That Scores 65 on the AA Decision Index
- 作者／機構:Asif Razzaq,MarkTechPost
- 連結:https://www.marktechpost.com/2026/10/15/liquid-ai-d1-ultra-decision-model/
#LiquidAI #DecisionModel #MultimodalAI #LLMAlternative #AIRouting #ContentModeration #MachineLearning #ArtificialAnalysis #AIInference #ModelPricing
原始資料 MarkTechPost · 收集於 2026-10-16
摘要原文
Liquid AI has released d1-ultra , its most capable decision model to date. The Liquid AI d1-ultra decision model reads text, images, or both. It answers typed questions with a probability for each possible answer. It does this in one forward pass and generates no output tokens. For developers, that means classification, routing and scoring calls without paying for generated text. It is available now on the Liquid API at $0.042 per million input tokens. A decision model answers bounded questions with probabilities instead of free text. Liquid’s decision model docs describe 3 question types: The docs recommend decision models for routing, moderation, guardrails, triage and LLM-as-judge replacement. They recommend a language model for summarization, chat and code generation. Each request carries a state (the context), optional images , and 1 or more typed questions . The model evaluates every question against the same input in 1 call. The d1 API reference shows responses with output_tokens always at 0. Liquid AI team states that d1-ultra uses a larger open-weight backbone than d1 , along with heavy Liquid-specific customizations. Combined with Liquid’s decision training pipeline, the gains are largest on knowledge tasks. Liquid AI team built d1-ultra for messy cases. Its examples include spotting a circuit board defect against a spec. Another is checking a damage claim against the customer’s photo. Liquid AI compared d1-ultra against its earlier d1 and TypeSafe’s Jev 1.14 across 112 evaluations. According to Liquid AI’s announcement , d1-ultra outperforms Jev 1.14 across all tasks at the same price. On the independent Artificial Analysis Decision Index , Liquid reports d1-ultra ranks #1 with a score of 65. The index averages 4 evaluations at 25% each: For context, Artificial Analysis’s Oct 13 launch article listed Jev 1.13 first at 58. GLiDE and Microsoft-Decision-1 followed at 54 each. d1-ultra runs on the Liquid API through the /decisions/v1/systemone endpoint with "model": "d1-ultra" . Liquid says Vercel AI Gateway and OpenRouter support is coming soon. You can test it first in the d1 Playground . Billing covers input tokens only. Per Liquid’s pricing notes , images cost 1.5 tokens per 32×32-pixel patch. A 1024×1024 image therefore costs 1,536 tokens. Each question is billed as its own prompt, including its text and all images. So 3 questions on 1 image bill that image 3 times. The endpoint docs cap requests at 8 images and 10,000 image patches. The full JSON body must stay under 4.5 MB. Check out the official blog post , the decision model docs , and try the d1 Playground . All credit goes to the researcher of this project. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter . Wait! are you on telegram? now you can join us on telegram as well. The post Liquid AI Releases d1-ultra: A Multimodal Decision Model That Scores 65 on the AA Decision Index appeared first on MarkTechPost .
由 claude-code/sonnet 自動生成