Liquid AI has released the Open d1 family, consisting of two open-weight multimodal models named d1-3B and d1-omni-600M. According to a report by MarkTechPost, the d1-3B model processes text and images, while the d1-omni-600M handles text with images or text with audio. A defining feature of these models is that they do not generate text output. Instead, each model delivers calibrated, typed answers in a single forward pass utilizing zero output tokens, aiming to facilitate real-time decision-making.
Understanding Zero Output Token Decision Models
Traditional language and multimodal models generate responses sequentially, predicting one token at a time until a complete sentence or paragraph is formed. By contrast, decision models designed to return structured answers without generating text outputs can theoretically reduce latency and computational overhead. This approach shifts the focus from generative conversational AI to direct classification and structured data extraction, which is critical for automated systems requiring instantaneous feedback loops. The open-weight release allows researchers and developers to deploy and evaluate this non-generative architecture locally.
Unknowns and What Lies Ahead
The announcement did not specify the exact hardware requirements needed to run either d1-3B or d1-omni-600M efficiently. Furthermore, specific benchmark scores detailing accuracy across different operational domains were omitted from the initial release details. Observers will need to monitor how these models perform in production environments compared to traditional generative architectures.
As developers begin testing the open-weight files, more data on their practical capabilities and integration limits will emerge.
Lead image: Jakub Pabis / Pexels
