Liquid AI Releases Open d1 Decision Models for the Edge

Liquid AI dropped two open-weight models on Hugging Face on Oct. 7, 2026, and they work nothing like the chatbots students are used to. The new d1-3B and experimental d1-omni-600M are decision models: you hand them a state — text, JSON, images or audio — along with a set of questions, and they return typed, calibrated answers in a single forward pass. No tokens generated, no output to parse, no prompt engineering required to get structured results.

That architectural choice is the whole story. Because the models skip autoregressive generation entirely, they’re fast in a way generative models simply can’t be at comparable sizes. On an NVIDIA Jetson AGX Thor — an edge device aimed at robotics and autonomous systems — d1-3B answers a question in 16 milliseconds. On the more affordable Jetson Orin Nano, it still comes in at 50 ms. On a desktop-class NVIDIA GeForce RTX 4090, latency drops to 8 ms. Liquid AI evaluated speed in collaboration with NVIDIA across all four of those platforms.

What Makes These Models Different

Both models are built on Liquid AI’s proprietary Liquid Foundation Models architecture, but they use different backbones. d1-3B is derived from LFM2.5-VL-3B, a decoder-only vision-language model that accepts text and images. The smaller d1-omni-600M starts from LFM2.5-Encoder-350M, a bidirectional encoder, and adds separate vision and audio encoders to handle all three modalities — making it capable of processing text paired with either images or audio. Liquid AI labels d1-omni-600M an early research release and cautions it is still under development.

On the Decision Index v0.2.1 — a public benchmark for this specific class of models — d1-3B scores 48.57, which puts it ahead of every sub-10B model tested and above Decider 35B-A3B (47.11), a model roughly 12 times its size. Across seven text-focused public datasets covering reading comprehension, toxicity detection, intent classification, medical question answering and cross-lingual understanding, d1-3B posts a mean score of 82.9. The 600M model reaches 78.4, outperforming Decider 2B despite having roughly a quarter of the parameters.

The wins are not uniform across every task. Decider 4B beats d1-3B on BoolQ, XNLI and the MASSIVE intent classification benchmark. And the audio capabilities of d1-omni-600M have no published benchmark to validate them — Liquid AI acknowledges that audio decision benchmarks remain an open problem in the field. All figures cited are company-reported and have not been independently reproduced.

Where This Fits in the Competitive Landscape

The broader category — open-weight, edge-optimized classifiers — already includes models like GLiNER2 (205M parameters, Apache 2.0 license), which handles extraction, classification and guardrails on text alone. What Liquid AI is adding is multimodal capability at this size class: d1-3B processes images alongside text, while d1-omni-600M adds audio, both while staying within latency budgets that matter for real-time systems. That combination is genuinely new at this footprint, even if the individual benchmarks invite scrutiny.

To demonstrate practical use, Liquid AI built ten live demos — gesture-controlled games, real-time content moderation and others — each running d1-3B in a loop over live camera input, one forward pass per frame. Those are available now in Liquid AI’s System One Arcade Hugging Face Space with no setup required.

Why Students and Early-Career Engineers Should Pay Attention

The barrier to building structured-decision AI applications just dropped considerably. Tasks like content moderation, intent routing, medical triage classification and visual inspection used to require either expensive API calls or custom classifier training from scratch. d1-3B’s documented use cases read like a shortlist of common capstone project and hackathon ideas: routing and triage, moderation, intent classification, extraction checks, reranking, LLM-as-a-judge scoring, agent guardrails and visual inspection.

Because the model produces zero output tokens, there’s no JSON parsing logic to write, no prompt template to maintain, and no runaway generation to debug when something goes wrong. For anyone who has spent time wrestling with getting a large language model to reliably return structured output, that reduction in integration complexity is meaningful.

The edge hardware angle matters for cost. A Jetson Orin Nano developer kit runs roughly $149–$199 — hardware a student can own — and d1-3B runs on it in 50 ms. That means real-time multimodal decision pipelines are now buildable without a cloud bill. Both models are free to download; Liquid AI has not announced API pricing for the open weights.

The model is available now on Hugging Face. Read the full release at Liquid AI’s announcement on Hugging Face.

Source: Hugging Face