Liquid AI Unveils d1: A Specialized Decision Model for Efficient AI-Powered Choices

Revolutionizing Structured Decision-Making with d1

The bigger takeaway is simple: In an era dominated by large language models (LLMs) capable of generating vast amounts of text, a new contender has emerged from Liquid AI, focusing not on creation, but on precision decision-making. Introducing d1, a groundbreaking decision model engineered specifically for structured choices, promising calibrated probabilities with zero output tokens.

Meanwhile, Many teams currently lean on general LLMs for tasks like classification, ticket routing, moderation, and scoring. While effective, these LLMs can be inefficient for such bounded problems. d1 steps in as a specialized, highly efficient alternative, designed to streamline these processes and deliver consistent, reliable outcomes.

What Makes d1 Different?

Unlike generative LLMs that compose new strings of text, d1 evaluates a given situation and returns a specific, typed answer from a predefined set of options. The core distinction lies in its fundamental design: it does not write text. This means every response from d1 results in usage.output_tokens being zero, leading to significant efficiencies.

In practical terms, Liquid AI’s guidance is clear: if your answer must be one of ‘N’ known options, a decision model like d1 is ideal. If the task requires the model to create novel text, then a traditional LLM remains the appropriate tool.

The Three Core Primitives of d1

d1 operates using three powerful primitives that can be combined in a single request, allowing the model to evaluate multiple questions against the same input state in one efficient round trip:

• Noul: This primitive handles simple yes/no questions, returning a probability between 0 and 1. For instance, determining if a customer message is a complaint might yield a 0.999 probability.
• Choice: Designed for scenarios where you need to select one option from a named set. It provides the top pick, the full probability distribution across options, and a confidence value. An example could be routing a double-charge ticket with 0.9997 confidence to ‘billing’.
• Score: This primitive rates input against an ordered rubric, returning a probability-weighted position. For an urgency rubric indexed from 0 to 3, a production outage might score 2.9995, indicating extreme severity.

How to Deploy and Interact with d1

For example, d1 is readily deployable today as a hosted API, accessible under the model name d1:free via the Liquid API. Notably d1 is currently API-only and not trainable, meaning there are no self-hostable weights (GGUF, MLX, or ONNX) available.

Interacting with the d1 API involves making a POST request to https://api.liquid.ai/decisions/v1/systemone. Each request comprises three key components:

1. The specific d1 model being used.
2. The ‘state,’ which can be plain text or a JSON object providing context.
3. The ‘questions,’ defined using the Noul, Choice, or Score primitives.

That said, Developers can integrate d1 into their applications using TypeSafe AI‘s SDKs, available for Python (typesafe-sdk) and TypeScript (@typesafe-ai/sdk).

Why Choose d1 Over General LLMs for Classification?

Migrating classification-type calls from general LLMs to d1 offers several compelling advantages:

• Zero Billed Output Tokens: Unlike LLMs which bill for output tokens even for single-word labels, d1 incurs no output token costs, leading to significant cost savings.
• Predictable Latency: Without a variable decoding loop, d1 offers consistent and predictable response times, crucial for real-time applications.
• No Schema Errors: d1’s answers always conform to the specified question type, eliminating malformed JSON outputs and the need for retries.
• Usable Uncertainty with Calibrated Probabilities: Instead of self-reported confidence scores, d1 provides calibrated probabilities, making thresholds practical. For example, a moderation system could block content above 0.8 probability, allow below 0.2, and flag the middle for human review.
• Fewer Round Trips: Multiple sequential classification tasks can be consolidated into a single d1 API call, reducing network overhead.
• Enhanced Consistency: Repeated evaluations of the same input yield more consistent results, minimizing “verdict flips” in automated systems.

Interestingly, These benefits highlight d1’s potential to dramatically improve the efficiency and reliability of automated decision-making processes.

When to Keep Your LLM

While d1 excels at structured decisions, it’s crucial to understand its boundaries. General LLMs remain indispensable for tasks requiring generative capabilities, such as:

• Summarization of lengthy texts.
• Drafting new content or emails.
• Engaging in multi-turn conversational chat.
• Generating code snippets.
• Performing complex, multi-step reasoning that involves synthesis and creation.

d1 in Action: The Road Decider Demo

However, Liquid AI showcases d1’s capabilities with a “Road Decider” cookbook, a pixel-art survival racer. In this demo, d1 uses a Choice question to determine the optimal navigation (left, center, or right) several times per second. A key learning from this demonstration emphasizes the importance of state design: providing d1 with per-lane summaries and distances to obstacles led to far more confident decisions compared to a raw grid representation of the road.

The Competitive Landscape

d1 enters a rapidly evolving category of non-generative decision models. While other models like TypeSafe Jev, Convai Laya, and AutoTrust JEV-27B offer similar primitives and zero output token billing, d1 distinguishes itself with its specific API-only access, robust primitive set, and focus on calibrated probabilities. The landscape is dynamic, with various offerings balancing access, pricing, and fine-tuning capabilities.

Key Takeaways for Developers and Businesses

Meanwhile, Liquid AI’s d1 decision model represents a significant step forward for applications requiring precise, efficient, and reliable structured decision-making. By offering calibrated probabilities with zero output tokens and a powerful set of primitives (Noul, Choice, Score), d1 is poised to optimize tasks like routing, moderation, triage, and reranking. For any scenario where the answer is one of ‘N’ known options, d1 offers a compelling, cost-effective, and high-performance alternative to general-purpose LLMs.

Expert Perspective

From an industry angle, the clearest signal around Liquid AI d1 decision model is how it may influence llms. The story reads less like a one-day spike and more like a marker of broader movement.

The next phase will depend on how quickly teams, regulators, or customers react. In practice, that gives Liquid AI d1 decision model room to reshape expectations across model over the near term.

For readers focused on practical impact, the best next step is to watch what changes around decision once attention turns into execution.

Frequently Asked Questions

Why does Liquid AI d1 decision model matter right now?

Revolutionizing Structured Decision-Making with d1The bigger takeaway is simple: In an era dominated by large language models (LLMs) capable of generating vast amounts of text, a new contender has emerged from Liquid AI, focusing not on creation, but on precision decision-making.

What broader change could Liquid AI d1 decision model signal?

Introducing d1, a groundbreaking decision model engineered specifically for structured choices, promising calibrated probabilities with zero output tokens.Meanwhile, Many teams currently lean on general LLMs for tasks like classification, ticket routing, moderation, and scoring.

What should the market watch next around Liquid AI d1 decision model?

While effective, these LLMs can be inefficient for such bounded problems.

Source: https://www.marktechpost.com/2026/09/29/liquid-ai-releases-d1-a-decision-model-that-returns-calibrated-probabilities-with-zero-output-tokens/