Liquid AI 发布多模态决策模型 d1-ultra，AA Decision Index 得分 65
Liquid AI Releases d1-ultra: A Multimodal Decision Model That Scores 65 on the AA Decision Index - MarkTechPost
Liquid AI 发布决策模型 d1-ultra，可读取文本、图像或两者，对每个问题返回各选项概率，一次前向传播完成且不生成输出 token。该模型已在 Liquid API 上线，定价为每百万输入 token 0.042 美元，仅按输入计费；Liquid 称其在 Artificial Analysis Decision Index 上以 65 分排名第一，并在其 112 项评测中领先全部 8 个类别。参数规模、上下文长度与权重均未披露，仅提供 API 访问，Vercel AI Gateway 和 OpenRouter 支持即将推出。
正文
当前显示原文，中文全文译文尚未就绪。
Add as a preferredsource on Google
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.
TL;DR
Size: Parameters and context length not disclosed. Built on a larger open-weight backbone than d1.
Runs on: Hosted Liquid API only. No downloadable weights or quantized variants announced.
Performance: Leads all 8 categories in Liquid’s 112-evaluation suite, per Liquid.
Best: 65 on the Artificial Analysis Decision Index, ranked #1 (Liquid-reported).
Worst: No public per-benchmark breakdown or parameter count yet; claims are vendor-reported.
What is a decision model?
A decision model answers bounded questions with probabilities instead of free text. Liquid’s decision model docs describe 3 question types:
Noul: a yes/no question that returns a probability between 0 and 1.
Choice: one label from a set, with a probability per label.
Score: a probability-weighted position on an ordered rubric.
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.
How does d1-ultra work?
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.
How does d1-ultra perform on benchmarks?
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:
AA-Omniscience-Decision: knowledge and hallucination
HLE-MultiChoice: reasoning
When2Call: tool calling
WildGuardTest: safety classification
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.
Comparison: d1-ultra vs other decision models
DeveloperLiquid AILiquid AITypeSafe AIMicrosoftFastino Labs
ParametersNot disclosedNot disclosedNot disclosedNot disclosedNot disclosed
InputText, imagesText, imagesTextNot disclosedNot disclosed
WeightsAPI onlyAPI onlyClosedNot disclosedNot disclosed
AA Decision Index65 (Liquid-reported)Not disclosed58 (v1.13, Oct 13)54 (Oct 13)54 (Oct 13)
Price$0.042 / 1M inputNot disclosed$0.042 / 1M input$31 / 1M tasks (AA)$119 / 1M tasks (AA)
Output tokens billedNoNoNoNot disclosedNot disclosed
How do you deploy d1-ultra, and what does it cost?
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.
Key Takeaways
d1-ultra returns probabilities for text and image decisions with 0 output tokens.
Liquid reports a #1 score of 65 on the Artificial Analysis Decision Index.
It costs $0.042 per 1M input tokens, the same rate as TypeSafe’s Jev.
Liquid claims the lowest calibration error in its tests, useful for routing thresholds.
Size, backbone and weights remain undisclosed; access is API-only.
Asif RazzaqWebsite | + postsBio ⮌
Asif RazzaqNVIDIA Releases Nemotron 3.5 Super VL, a 120B Open Multimodal Agent Model
Asif RazzaqTypeSafe AI Releases Jev 1.14 With Images, Box Type and 262k Context
RELATED ARTICLESMORE FROM AUTHOR
NVIDIA Releases Nemotron 3.5 Super VL, a 120B Open Multimodal Agent Model
openKylin Outlines Its Vision for Reimagining Linux in the AI Era at Open Source Summit Europe 2026
TypeSafe AI Releases Jev 1.14 With Images, Box Type and 262k Context
Meet Grok Voice 3.0: SpaceXAI Voice Agent Model Ranks #1 on Artificial Analysis