Liquid AI Releases Two Open Decision Models for Edge

Liquid AI open decision models release

Liquid AI just dropped two open decision models built for edge hardware, marking the first public look at its d1 family. The announcement landed October 7, 2026, covering d1-3B and d1-omni-600M. The pitch is straightforward: these are alternatives to generative models for when you need a single, structured answer-think a yes/no, a category label, or a numeric score-instead of a stream of tokens. Both are up on Hugging Face under open weights right now, and there’s a demo space called System One Arcade so you can kick the tires without spinning up local infra.

Unlike the LLMs hogging the spotlight, decision models crank out one forward pass per query. That design kills the latency of token-by-token generation, making them practical for real-time control loops, on-device classification, or anywhere you need an answer in milliseconds. Liquid AI says d1-3B hits a Decision Index 0.2.1 score of 48.57, claiming the top spot among models under ten billion parameters and edging out the larger Decider 35B-A3B at 47.11. The benchmark suite spans seven public datasets covering reading comprehension, toxicity detection, intent classification, medical QA, and cross-lingual understanding. On that suite, d1-3B posts a mean of 82.9, while the smaller d1-omni-600M reaches 78.4, beating Decider 2B's 77.1 despite using roughly a quarter of the parameters.

Decision model architectures compared

The two models come from very different architectural roots. d1-3B is distilled from LFM2.5-VL-3B, a decoder-only vision-language model that takes text and images. d1-omni-600M starts from LFM2.5-Encoder-350M, a bidirectional encoder Liquid AI extended with separate vision and audio encoders to handle text-plus-image or text-plus-audio pairs. The company labels d1-omni-600M an early research release and notes that audio decision benchmarks are still an open problem, so they aren't reporting audio-specific numbers yet.

Edge inference speed benchmarks

Speed figures published with NVIDIA show the edge focus isn't just marketing fluff. On a Jetson AGX Thor, d1-3B answers a single question in 16 milliseconds; on a Jetson AGX Orin 64 GB it’s 26 ms, and on a Jetson Orin Nano it’s 50 ms. Batching three questions together costs only a 1.3× multiplier on the Thor, totaling 20 ms. Desktop GPUs go even faster: an RTX 4090 returns an answer in 8 ms and an AMD MI325X in 9 ms. Image inputs add modest overhead-17 ms on the 4090 and 18 ms on the MI325X for a 384-pixel image. Liquid AI hasn't published speed data for d1-omni-600M yet since it’s still early days.

For developers wanting to integrate today, the requirements are straightforward. You’ll need transformers 5.14 or newer, plus PyTorch, torchvision, and Pillow. Because the models ship custom modeling code, loading them requires `trust_remote_code=True`. The API centers on a `system_one` method that accepts a state, text, an image, or both, plus a dictionary of named questions. Each question specifies a type like `noul` for yes/no, `choice` for categorical selection with labeled criteria, or `score` for an ordinal rating. A single call can evaluate several questions over the same state, and a batched variant `system_one_batch` processes multiple states with no padding overhead. The model card for d1-omni-600M includes separate usage notes for its dual-modality input paths.

Local AI decision use cases

The practical upshot for AI builders is that structured decision tasks-routing a support ticket, flagging toxic content, validating a medical triage answer, counting objects in a camera frame-can now run locally on hardware costing a few hundred bucks that sips single-digit watts. That unlocks product categories previously blocked by cloud latency, connectivity needs, or privacy constraints. Teams already on Hugging Face can drop these into existing pipelines with minimal friction, and the System One Arcade space gives you a low-risk way to validate behavior on real data before committing engineering cycles.

Anyone evaluating this should start by running the arcade demos against their own edge-case inputs. If quality holds, the next step is benchmarking on your target hardware with the actual batch sizes and input modalities production will see. Since d1-omni-600M is explicitly labeled experimental, treat it as a research prototype and plan for API changes; d1-3B looks ready for serious evaluation in latency-sensitive pipelines. Both are open-weight, so fine-tuning or distillation stays on the table if out-of-the-box performance falls short on a niche task.

Source: Hugging Face