AI News Feed
Market watch
BMW details 2027 i3: 468-mile range, $62,850 starting priceTranslation Earbuds Work in Three Steps, but Speed and Accuracy Still VaryRainbow Six Tactics Director Says 15 Operators Built to Avoid Overlap, Tachanka and Lion Miss the CutRussia’s USC unveils underwater robotic systems for oil, gas and ship repairxAI's dot.com Domain Redirects to Grok After OpenAI Unveils 'Dots' AgentStudy Finds Connected Cars and Their Apps Funnel Driver Data to Ad and Tracking Firms
BMW details 2027 i3: 468-mile range, $62,850 starting priceTranslation Earbuds Work in Three Steps, but Speed and Accuracy Still VaryRainbow Six Tactics Director Says 15 Operators Built to Avoid Overlap, Tachanka and Lion Miss the CutRussia’s USC unveils underwater robotic systems for oil, gas and ship repairxAI's dot.com Domain Redirects to Grok After OpenAI Unveils 'Dots' AgentStudy Finds Connected Cars and Their Apps Funnel Driver Data to Ad and Tracking Firms
Companies

Liquid AI Releases d1, a Decision Model That Returns Calibrated Probabilities With Zero Output Tokens

Liquid AI has launched d1, a hosted decision model that answers typed questions with calibrated probabilities instead of text, targeting classification, routing, scoring and moderation work that teams currently send to general-purpose LLMs.

AI News Feed · 2026-09-30 07:38

The model is aimed at work many teams still send to general-purpose large language models: classification, ticket routing, scoring, moderation, reranking and LLM-as-judge checks. According to a MarkTechPost report published on September 29, 2026, callers supply context and a set of typed questions, and d1 returns probabilities for each defined outcome in a single call. Liquid AI's model library lists d1 as API only and not trainable, so there are no GGUF, MLX or ONNX weights for teams that want to self-host it.

A decision model, as Liquid AI describes it, evaluates a situation and returns a typed answer drawn from options defined before the call. It does not write text, and in every response usage.output_tokens is 0. The company's migration guide states a simple rule: if the answer is one of N known options, use a decision model; if the model must compose a new string, keep the LLM.

Three question types are supported. Noul poses a yes or no question and returns a probability between 0 and 1; in Liquid AI's example, the question "Is this message a complaint?" returned 0.999. Choice selects one option from a named set and returns the top pick, the full distribution and a confidence value; a double-charge ticket scored 0.9997 on "billing." Score rates input against an ordered rubric and returns a probability-weighted position, with levels indexed from 0 so that a four-level urgency rubric spans 0 to 3; a production outage scored 2.9995. The three types can be mixed in one request, and the model evaluates every question against the same state in one round trip.

Each request contains the model, the state as plain text or a JSON object, and the questions. Calls go to POST https://api.liquid.ai/decisions/v1/systemone. API keys are issued from console.liquid.ai and begin with liquid_. The supported clients are TypeSafe AI's typesafe-sdk for Python and @typesafe-ai/sdk for TypeScript.

Liquid AI's migration guide lists concrete differences from an LLM with structured output: no billed output tokens, since an LLM bills output even for a one-word label; predictable latency, because no decoding loop grows with output length; no schema errors, since answers always match the question type; calibrated probabilities rather than a self-reported number; and fewer round trips, turning three sequential classification calls into one. The company's moderation example blocks above 0.8, allows below 0.2 and sends the middle band to human review, while its routing example falls back to the most capable model tier when router confidence drops below 0.5. Liquid AI also says repeated evaluations of the same input are more consistent, which it says reduces verdict flips. Summarization, drafting, multi-turn chat, code generation and complex multi-step reasoning remain LLM work.

A published demo, the road-decider cookbook, is a pixel-art survival racer in which d1 uses a Choice question to pick left, center or right on every decision tick, roughly two to five times per second depending on game speed. The app is written in vanilla JavaScript on Node.js 18 or later, with a Vite proxy that keeps the API key server-side, and a "Jev vs d1" mode races d1 against TypeSafe's typesafe/jev-1.13 through OpenRouter. The stated lesson concerns state design: per-lane summaries with distance to the first obstacle produced more confident decisions than a raw grid of the road.

d1 enters a small category of non-generative decision models. In a feature comparison checked on September 29, 2026, TypeSafe's Jev 1.13 offers the same Noul, Choice and Score primitives through OpenRouter at $0.042 per 1 million input tokens with a 32K context window and 0 output tokens; Convai's Laya publishes 421 million parameters for English and 322 million for multilingual use, context of 512 or 1,024 tokens, 0 output tokens and open Apache 2.0 weights; AutoTrust's JEV-27B uses a 27 billion parameter backbone with 108.9 million trainable parameters, supports true/false, Choice from 2 to 16 options and Score from 0 to 5, produces 0 output tokens for decisions while retaining text generation, and is self-hostable under Apache 2.0 with LoRA fine-tuning. d1's context window, model size, paid pricing and fine-tuning support are not published; Convai and AutoTrust are the only two in the comparison released with open weights.