Liquid AI’s Open d1 Explained: AI Models That Make Decisions

Episode Overview

• This episode explores the rapid rise of "decision models," focusing on Liquid AI's newly released open-weight model family, Open D1 (featuring d1-3B and d1-omni-600M).
• It highlights the architectural shift from traditional generative language models to highly efficient decision models that bypass token-by-token generation to return structured actions, probabilities, and scores.
• The episode details how Liquid AI uses a hybrid architecture (combining convolutions and attention) to enable real-time, multimodal (text, vision, and audio) reasoning directly on edge devices.
• It provides a framework for understanding how developers and roboticists can utilize these small, fast models to orchestrate complex physical systems like autonomous labs and robots.

Key Concepts

• Generative vs. Decision-Centric AI: Traditional LLMs generate text token-by-token, which is slow and expensive for simple tasks. Decision models process input in a single forward pass and directly return constrained judgments (e.g., multiple choice, yes/no probabilities, or numeric scores), massively reducing computation cost and latency.
• Multimodality on the Edge: While early decision models like Jev were text-only, Open D1 extends this paradigm to vision and audio. This allows physical devices to make instantaneous decisions based on real-world sensory inputs without relying on cloud-based APIs.
• Convolutions Combined with Attention: To bypass the heavy computational demands of standard transformer attention, Liquid AI's LFM 2.5 architecture mixes short convolutions (to quickly parse local neighboring tokens) with attention mechanisms (to link global context). This hybrid setup is optimized to run efficiently on embedded CPUs.
• The "Edge" Computing Paradigm: Running models locally on edge devices (like NVIDIA Jetson or smart assistants) removes network latency, ensures continuous operation during internet dropouts, and protects user privacy by keeping sensitive raw audio and video data on-device.

Quotes

• At 1:21 - "They can work with images, and one of them also accepts audio as an experiment, and they are open-weight—I love that." - Explaining the shift of decision models into multi-modality and open-source accessibility.
• At 4:30 - "How much of this work can a model do without generating a response?" - Identifying the core question behind decision models, showcasing why bypassing generative text makes workflows cheaper and faster.
• At 11:06 - "The ultimate moat is fast and deep integration." - Relaying Ramin Hasani's (CEO of Liquid AI) vision that AI competition is shifting away from raw parameter size toward how cleanly models integrate into real-world software and hardware.

Takeaways

• Deploy Open D1 models directly onto low-power edge hardware (such as Nvidia Jetson or Apple M-series chips) to run ultra-fast, local vision and audio classifications without cloud latency.
• Replace generative LLM steps in your automation pipelines with decision models if your system only requires structured routing outputs (like categorizing customer support tickets or verifying state checks), saving on API costs and generation time.
• Build robust state-tracking systems for physical automation or autonomous labs by passing natural-language questions (e.g., "Is the test tube cap sealed?") to a local visual decision model, feeding the yes/no probability directly into your system's control loop.