Llull

Introduction

Introducing Llull, our fast model for structured knowledge and the relationships between concepts. It scores 86% on our Structured Analysis evaluation, responds in an average of 2.4 seconds, and is the model behind Tiff Lite.

Llull is named after Ramon Llull, whose work explored systematic ways of combining concepts to investigate questions. That idea is the foundation of the model: intelligence becomes more powerful when knowledge can be connected.

Llull’s results on our four evaluations:

• Contextual Reasoning. 84%. This evaluation measures interpreting context, resolving ambiguity and reaching the correct conclusion.
• Knowledge Reasoning. 81%. This evaluation measures multi-step reasoning across knowledge-intensive questions.
• Structured Analysis. 86%. This evaluation measures organizing information, identifying relationships and producing logically structured answers.
• Complex Problem-Solving. 79%. This evaluation measures multi-step problems that require reasoning across several pieces of information.

Aurelius, our newer model, scores 7 to 9 points higher on all four evaluations, with an average first response about 17% slower. Llull is the faster of the two.

What Llull is built for

• knowledge-intensive applications
• structured analysis and structured question answering
• connecting information across domains
• research and information synthesis
• classification and organization of knowledge

The problem Llull works on

Traditional information systems are good at finding individual pieces of information. The harder problem is understanding how those pieces relate. A question about an economy may involve policy. A question about culture may involve history and language. A question about governance may involve institutions, people and events. Llull is our exploration of models that can work with that network.

Performance and speed

Llull’s scores on the same four evaluations, shown next to Aurelius.

All Aurelius and Llull results come from Trendify Labs internal evaluations. Both models were run on the same evaluation sets, with the same prompts and inference conditions. We have not yet published results on third-party benchmarks, and we will add broader comparisons as our evaluation suite expands.

Speed

Llull is the faster of our two models, with an average first response of 2.4 seconds compared to Aurelius’s 2.8 seconds.

Safety

We evaluated both models against a structured internal safety suite covering harmful requests, privacy-sensitive requests, misinformation, prompt injection, unsafe instructions, sensitive personal data and adversarial prompts. Tests ran under the same inference and system-safety conditions for both models.

Safeguards

• system-level safety policies
• input and output filtering
• instruction hierarchy
• refusal handling
• prompt-injection resistance
• continuous evaluation against internal adversarial test sets

Known limits

Neither model is infallible. Both can produce incorrect or incomplete information, particularly when prompts are ambiguous or the information is unavailable. Llull and Aurelius were tested under the same conditions; Aurelius can spend more time reasoning on difficult requests, resulting in higher latency.

Getting started

Llull powers Tiff Lite, and Aurelius powers Tiff Normal. Both models are available through the Tiff platform, including the Tiff app, the web app, Telegram and Instagram.

Developers can use Llull through the API with the model name llull. To get API access, contact Trendify Labs.

This page will be updated as we publish more results.