Decision models: when your AI should return a probability, not a paragraph

What happened

Liquid AI released two open-weight decision models, d1-3B and d1-omni-600M, which are designed to make decisions by returning probabilities instead of text outputs, offering faster processing and lower costs compared to traditional chat models.

Why it matters

These models can improve efficiency in AI systems by reducing latency and cost for tasks that require quick, probabilistic decisions, such as ticket classification and moderation.

• Adopt decision models for tasks requiring quick, probabilistic decisions to reduce costs and latency.
• Use confidence scores to route tasks efficiently, automating confident decisions and escalating uncertain ones.
• Evaluate the models on your specific use cases and check calibration before deployment.

The bigger picture

The models can lead to cost and latency savings in AI systems by efficiently handling high-volume, low-complexity decisions, allowing larger models to focus on more complex tasks.

Liquid AI's decision models offer a new approach to AI decision-making, providing faster, cheaper, and more efficient solutions for tasks that require probabilistic outputs rather than text generation.

Technical details

• d1-3B: 3.1B parameters, trained from LFM2.5-VL-3B vision-language backbone, handles text and images.
• d1-omni-600M: 587M parameters, trained from a bidirectional encoder with added vision and audio encoders, handles text plus image or audio.
• Decision models return typed answers with probabilities, such as yes/no, choice among options, or scores on a scale.
• They use a single forward pass to generate answers, significantly reducing processing time compared to chat models.

FAQ