Liquid AI Launches d1-ultra Multimodal Decision Model for Classification

Liquid AI has released d1-ultra, its most advanced decision model, which provides calibrated probability scores for text and image inputs without generating output tokens. According to MarkTechPost, the model is available immediately on the Liquid API at a cost of $0.042 per million input tokens. The system is designed to answer bounded questions with precise likelihoods rather than free-form text, enabling efficient classification and routing tasks for developers.

The d1-ultra model operates by evaluating typed questions against input context in a single forward pass. This architecture eliminates output token costs, as the model returns zero output tokens per request. According to the vendor, the model supports three specific question formats: yes/no queries returning a probability between zero and one, choice questions selecting one label from a set with associated probabilities, and score questions determining a position on an ordered rubric.

Technical Specifications and Performance

Liquid AI states that d1-ultra is built on a larger open-weight backbone than its predecessor, d1, combined with proprietary customizations. The company claims the model leads all eight categories in its internal 112-evaluation suite. According to Liquid AI, d1-ultra outperforms TypeSafe’s Jev 1.14 across all tasks at the same price point. The model’s performance gains are reported to be most significant on knowledge-intensive tasks due to its specialized decision training pipeline.

On the independent Artificial Analysis Decision Index, Liquid reports that d1-ultra ranks number one with a total score of 65. This index averages four evaluations at twenty-five percent weight each, including a measure for knowledge and hallucination known as AA-Omniscience-Decision. However, the company has not disclosed the model’s parameter count, context length, or a public per-benchmark breakdown of its performance metrics.

Use Cases and Limitations

The Liquid AI team recommends d1-ultra for specific operational tasks such as routing, moderation, guardrails, triage, and replacing LLM-as-judge workflows. The model is intended for scenarios requiring structured decisions rather than content generation. According to the vendor, typical use cases include identifying circuit board defects against technical specifications or verifying insurance damage claims using customer-submitted photos. For tasks like summarization, chat, or code generation, the company advises using traditional language models instead.

Access to d1-ultra is restricted to the hosted Liquid API. No downloadable weights or quantized variants have been announced for local deployment. While the model accepts optional images alongside text inputs, its core function remains the evaluation of multiple typed questions against a shared input state. The absence of output token generation means developers pay only for the input data processed by the model.

What should you do?

Security and technology professionals should evaluate whether their current classification or routing workflows could benefit from a probability-based model that eliminates output token costs. Teams should review the vendor’s documentation to understand the specific question types supported by the d1 API. Organizations should also compare the reported performance metrics against their internal requirements for accuracy and hallucination resistance.

What to do and how to stay safe: Liquid AI

• Review the Liquid API documentation to understand the specific formats for yes/no, choice, and score question types.
• Test the model with your own text and image data to verify its accuracy on your specific classification tasks.
• Compare the cost structure of $0.042 per million input tokens against your current language model expenses for routing tasks.
• Monitor for any future updates regarding parameter counts or downloadable weights if local deployment is required.

Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation

General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

What is the cost of using Liquid AI’s d1-ultra model?

The model costs $0.042 per million input tokens, with no charge for output tokens since the model returns zero output tokens.

Can I download the d1-ultra model weights for local use?

No, the model is available only through the hosted Liquid API, with no downloadable weights or quantized variants announced.

How does d1-ultra differ from standard language models?

d1-ultra returns calibrated probabilities for bounded questions in a single forward pass without generating free-form text or output tokens.