Product Introduction
- Definition: Liquid d1.2 and d1.2-ultra are state-of-the-art omnimodal decision models developed by Liquid AI. Operating within the artificial intelligence and machine learning landscape, these foundation models process multi-modal inputs—including text, images, and audio—to execute complex extraction, classification, and reasoning workloads while providing statistically calibrated probabilities for every output.
- Core Value Proposition: These decision models exist to bridge the gap between automated data processing and human oversight by outputting precise, calibrated confidence scores. This core capability allows enterprise systems to automatically route low-confidence predictions to human reviewers, drastically reducing error rates in high-stakes environments. By integrating audio inputs and introducing advanced document reading, object localization, text span detection, numerical estimation, and multi-select classification, Liquid d1.2-ultra delivers industry-leading accuracy on the Artificial Analysis Decision Index v2.1 while maintaining high compute efficiency and cost-effectiveness.
Main Features
- Omnimodal Input Processing & Audio Support: Liquid d1.2 integrates native handling for text, images, and audio modalities within a single decision-making endpoint. Audio processing capabilities allow models to ingest spoken queries, evaluate acoustic environments (such as identifying background sirens), or classify spoken languages across 43 distinct dialects (via datasets like FLEURS) with 95% to 97% accuracy.
- Advanced Decision Tasks & Structured Extraction: The models introduce five core production tasks: reading precise values from structured documents like invoices (returning exact character matches or a "not found" state to prevent hallucination); identifying text spans like personal data (PII) with character positions; localizing objects in images via bounding boxes; estimating numerical quantities with specific error ranges; and executing multi-select classifications with per-topic probabilities.
- Calibrated Confidence Scoring: Designed for enterprise risk management, the models output statistically robust probability distributions. With a calibration error of just 0.8 points for d1.2-ultra (tested across 13 benchmarks and 22,505 questions), the model's predicted confidence closely matches its actual accuracy, enabling developers to build automated threshold-based workflows that flag uncertain predictions.
Problems Solved
- Pain Point: Generative AI models traditionally suffer from hallucination and uncalibrated overconfidence, making them unreliable for automated document processing, compliance checks, and data extraction where incorrect guesses lead to costly operational errors.
- Target Audience: Enterprise AI Engineers, Machine Learning Developers, Data Science Teams, and Automation Architects building production pipelines in finance, automotive, e-commerce, and healthcare.
- Use Cases: Automated accounts payable processing (extracting line items and totals from invoices with fallback routing for ambiguous fields), compliance auditing (identifying and redacting PII spans in legal or medical documents), visual quality inspection (locating components and counting objects in manufacturing photos), and multi-label categorization of customer support tickets.
Unique Advantages
- Strengths & Limitations (Pros & Cons):
- Pros: Industry-leading performance on the Artificial Analysis Decision Index v2.1 (scoring 75 with d1.2-ultra); superior probability calibration compared to competitor alternatives; highly cost-efficient token billing architecture ($0.02 for d1.2 and $0.042 per million input tokens); comprehensive omnimodal input support (text, vision, audio); native "not found" handling to eliminate hallucinations.
- Cons: Availability is currently restricted primarily to the Liquid API console (with OpenRouter and Vercel integrations coming soon); fine-tuning ecosystem and public documentation are still maturing relative to legacy foundation model providers.
- Key Alternatives & Differentiation: Compared to generalized large language models like GPT-6 Luna (Decisions API) or specialized extraction models like Jev 1.14 and Perplexity Decider v1.3, Liquid d1.2-ultra demonstrates significantly lower calibration error (0.8 points vs. up to 7.2 points for competitors). Furthermore, while standard LLMs require complex prompt engineering to return structured bounding boxes, numerical ranges, or confidence metrics, Liquid's decision models natively output these data structures with deterministic formatting and superior cost-to-performance ratios.
Frequently Asked Questions (FAQ)
- What is the difference between Liquid d1.2 and d1.2-ultra? Liquid d1.2-ultra is the flagship model optimized for maximum accuracy, achieving a top score of 75 on the Artificial Analysis Decision Index v2.1, whereas d1.2 offers a more cost-effective entry point ($0.02 per million input tokens versus $0.042 for ultra) while retaining the same comprehensive set of omnimodal features and five new decision tasks.
- How does probability calibration work in Liquid decision models? Calibration ensures that when the model assigns a specific percentage score to an answer (e.g., 80% confidence), it is statistically correct approximately 80% of the time. This tight alignment allows developers to establish automated threshold rules to route low-confidence outputs directly to human-in-the-loop review queues.
- How are images and audio billed on the Liquid API? Text and images are billed based on input tokens, where an image costs 1.5 tokens per 32x32-pixel patch at the standard text rate (e.g., a 1024x1024 image consumes 1,536 tokens). Each request bills its base inputs once, and additional questions appended to the payload add only their specific text token overhead.
- Where can I access Liquid d1.2 and d1.2-ultra?
Both models are currently available directly on the Liquid API via the Liquid Console (console.liquid.ai) using dedicated endpoint routing (
https://api.liquid.ai/decisions/v1/systemone), with integration support for OpenRouter and Vercel rolling out soon.
