Liquid AI Released Open-Weight Decision Models
New multimodal models bypass text generation to enable real-time routing and intent classification at low latency.
Updated on Oct. 7, 2026 in Artificial Intelligence
Live Poll
Would you use AI tools designed for direct decision-making rather than generative text responses?
Liquid AI has released the d1-3B and d1-omni-600M models, two open-weight architectures designed to perform real-time decision tasks without generating text output. These models, which support image and audio inputs, are now available for commercial use under specific revenue constraints.
Why it matters
By returning probability distributions instead of generating text, these models prioritize speed and efficiency for critical backend tasks like content moderation and automated routing. This shift signals a move toward specialized, high-velocity inference architectures in commercial applications.
The d1-3B model, which features 3.12 billion parameters, achieves a request latency of 8 milliseconds when running on an RTX 4090 GPU. The architecture utilizes an averaged combination of the LFM2.5-2.6B model and a vision-language model to process inputs.
The players
Liquid AI
A developer of alternative neural network architectures focusing on high-efficiency, non-autoregressive inference.
Hugging Face
A primary hub for open-source machine learning models and datasets where the d1 checkpoints are hosted.
The details
Rather than producing text tokens—the standard unit of data processed by LLMs—the d1 models return probability distributions over fixed options in a single forward pass. This method eliminates the computational overhead of autoregressive text generation, which requires the model to predict one word at a time. The d1-3B model accepts text and image inputs, while the d1-omni-600M architecture supports text combined with images or audio clips up to 30 seconds long.
Timeline
October 7, 2026: Liquid AI released the d1 decision models.
The Tech Race
This release follows the development trajectory of the LFM2.5 model series, focusing on efficient, specialized inference rather than general-purpose text generation. It represents a direct challenge to standard LLM-based workflows that rely on costly and slow token-by-token generation for basic routing tasks.
Developers and startups with under $10 million in annual revenue can use the d1 models for free under the LFM Open License v1.0. These models are intended to replace slower text-generation workflows for applications like intent classification, moderation, and automated routing.
The takeaway
Liquid AI is trading conversational fluency for raw decision-making speed in enterprise environments. Watch for industry adoption benchmarks to see if these models can effectively displace traditional LLMs in production routing pipelines.
Further reading
For broader trends in inference efficiency, see the Artificial Intelligence section.
Live Poll
Would you use AI tools designed for direct decision-making rather than generative text responses?
