Liquid AI d1-ultra is a new multimodal decision model that processes text and images and returns probabilities for answers. The model evaluates multiple typed questions in one call and generates no output tokens. According to Liquid AI, d1-ultra ranks #1 with a 65 on the Artificial Analysis Decision Index. On the Liquid API, pricing is $0.042 per 1M input tokens, with no charges for generated text.
What is Liquid AI d1-ultra?
d1-ultra is Liquid AI’s most capable decision model to date. It reads text, images, or both and returns probabilities for each possible answer instead of free-form text, with 0 output tokens.
Liquid’s docs describe three question types: Noul (yes/no with probability 0–1), Choice (one label from a set with per-label probabilities), and Score (a probability-weighted position on an ordered rubric). Recommended uses include routing, moderation, guardrails, triage, and LLM-as-judge replacement. For summarization, chat, and code generation, Liquid suggests language models.
For business readers: you ask typed, bounded questions and get calibrated probabilities that fit neatly into rules and thresholds without parsing long answers.
How does d1-ultra work?
Each request carries state (context), optional images, and one or more typed questions. The model evaluates all questions against the same input in a single call. In the d1 API, output_tokens are always 0, so you don’t pay for generated text.
The Liquid AI team states d1-ultra uses a larger open-weight backbone than d1 with significant Liquid-specific customizations. Combined with Liquid’s decision training pipeline, the biggest gains appear on knowledge tasks. The model targets messy real-world cases: examples include spotting a circuit board defect against a spec and checking a damage claim against a customer photo.
How does d1-ultra perform on benchmarks?
Per Liquid, d1-ultra leads all 8 categories in a suite of 112 evaluations. On the independent Artificial Analysis Decision Index, Liquid reports d1-ultra at 65 and #1 overall.
The index averages four evaluations at 25% each: AA-Omniscience-Decision (knowledge and hallucination), HLE-MultiChoice (reasoning), When2Call (tool calling), and WildGuardTest (safety classification). For context, Artificial Analysis’s Oct 13 launch article listed Jev 1.13 first at 58, with GLiDE and Microsoft-Decision-1 following at 54 each.
How to run d1-ultra on the Liquid API?
d1-ultra runs via the Liquid API at /decisions/v1/systemone with "model": "d1-ultra". It’s available as a hosted API only (no downloadable weights). You can first test it in the d1 Playground. Liquid also says Vercel AI Gateway and OpenRouter support is coming soon.
Billing covers input tokens only at $0.042/1M. Images cost 1.5 tokens per 32×32-pixel patch; thus a 1024×1024 image costs 1,536 tokens. Each question is billed as its own prompt, including its text and all images — so 3 questions over 1 image bill the image three times. Endpoint limits: up to 8 images and 10,000 patches; the full JSON body must be under 4.5 MB.
How does d1-ultra compare to alternatives?
Liquid compared d1-ultra to its prior d1 and TypeSafe Jev 1.14 across 112 evaluations. According to Liquid, d1-ultra outperforms Jev 1.14 across all tasks at the same price. Here are the key attributes and pricing from the available data.
| Feature | d1-ultra | d1 | Jev (TypeSafe) | Microsoft-Decision-1 | GLiDE (Fastino) |
|---|---|---|---|---|---|
| Developer | Liquid AI | Liquid AI | TypeSafe AI | Microsoft | Fastino Labs |
| Parameters | Not disclosed | Not disclosed | Not disclosed | Not disclosed | Not disclosed |
| Input | Text, images | Text, images | Text | Not disclosed | Not disclosed |
| Weights | API only | API only | Closed | Not disclosed | Not disclosed |
| AA Decision Index | 65 (Liquid-reported) | Not disclosed | 58 (v1.13, Oct 13) | 54 (Oct 13) | 54 (Oct 13) |
| Price | $0.042 / 1M input | Not disclosed | $0.042 / 1M input | $31 / 1M tasks (AA) | $119 / 1M tasks (AA) |
| Output tokens billed | No | No | No | Not disclosed | Not disclosed |
Часті питання
What is a decision model and how does it work?
A decision model returns probabilities for bounded questions instead of free text. Liquid defines three types: Noul (yes/no with 0–1), Choice (a label from a set with per-label probabilities), and Score (a position on an ordered rubric). d1-ultra reads text and/or images and evaluates multiple questions in one call. In the d1 API, output_tokens = 0, so you pay only for input.
d1-ultra vs Jev comparison?
Liquid’s announcement says d1-ultra outperforms TypeSafe Jev 1.14 across 112 evaluations at the same $0.042 / 1M input rate. On the Artificial Analysis Decision Index, d1-ultra reports 65, while the Oct 13 article listed Jev 1.13 at 58. GLiDE and Microsoft-Decision-1 followed at 54. Parameters and detailed breakdown aren’t public.
How much does d1-ultra cost (pricing)?
The rate is $0.042 per 1M input tokens. Output tokens aren’t billed. Images are charged at 1.5 tokens per 32×32-patch, so a 1024×1024 image costs 1,536 tokens. Each question is billed as its own prompt and includes all images attached. Three questions over one image bill that image three times.
d1-ultra for content moderation?
Liquid recommends decision models for moderation, guardrails, routing, triage, and as an LLM-as-judge replacement. d1-ultra returns calibrated probabilities, which fit safety thresholds and policy checks. Liquid’s examples include checking a circuit board defect and validating a damage claim against a photo — a transferable approach to content moderation with rule-based thresholds.
Based on MarkTechPost.
