Edge-Deployed Multimodal Open D1 Decision Models

Meta description: Open d1 decision models for edge devices score 48.57 on Decision Index 0.2.1 under 10B, with d1-3B leading and fast multimodal answers on GPUs.

Tags: d1-3B, d1-omni-600M, edge AI, Decision Index, Hugging Face


The team behind Liquid Foundation Models has made open d1 decision models available as open-weight models on Hugging Face today, and the release gives builders a new option for small, fast, local decision-making. The headline result is d1-3B: it scores 48.57 on the Decision Index 0.2.1, placing it as the best decision model under 10B in that comparison. It sits ahead of every 4B and 9B model tested and also ahead of Decider 35B-A3B, which scores 47.11. Alongside d1-3B, the release includes d1-omni-600M, an early research release designed to work with text plus image inputs or text plus audio inputs.

Decision models answer differently from chat models

The d1 systems are decision models built on Liquid Foundation Models, shortened as LFMs. They are not intended to generate long text token by token like a chatbot. Instead, they produce an answer through a single forward pass. That difference matters for edge devices because a one-pass response can be much easier to time, much easier to deploy, and much easier to use inside products that need a quick classification, judgment, or routing decision rather than a written reply.

The two released models take different paths. d1-3B is trained from LFM2.5-VL-3B, the team’s latest vision-language model, or VLM, with a decoder-only architecture. It accepts text and images as inputs. d1-omni-600M is trained from LFM2.5-Encoder-350M, a bidirectional encoder, and adds vision and audio encoders so it can handle all three modalities. In practice, d1-omni-600M accepts either text and image, or text and audio, as inputs. It is currently in an early research release and is undergoing further development.

Benchmark results across seven public datasets

The release reports results from seven public datasets spanning reading comprehension, toxicity detection, intent classification, medical QA, and cross-lingual understanding. d1-3B achieves a mean score of 82.9, which is the highest score in the table and is above Decider 4B. d1-omni-600M scores 78.4, surpassing Decider 2B at 77.1 while using only a quarter of the parameters.

The team also states that d1-3B retains the vision capabilities of its LFM2.5-VL-3B backbone on standard vision benchmarks, and that d1-omni-600M handles all three modalities. However, the release does not report vision or audio benchmark numbers. The reason given is that Decision Index v0.3 includes only a private vision split, and audio decision benchmarks are currently an open problem.

Edge speed on NVIDIA hardware

Speed is the other major part of the release. In collaboration with NVIDIA, d1-3B was evaluated on the NVIDIA stack across NVIDIA GeForce RTX 4090, NVIDIA Jetson AGX Thor, Jetson AGX Orin 64 GB, and Jetson Orin Nano. For the Jetson line, d1-3B answers a question in 16 ms on a Jetson AGX Thor, 26 ms on a Jetson AGX Orin, and 50ms on a Jetson Orin Nano. Because d1-omni-600M is an early research release, no speed numbers for it are reported in this release.

The edge inference story is simple: d1-3B answers a single question in under 50 ms on every measured device. The scaling behavior is also useful. Three questions take only 1.3x the time of one, with the Jetson AGX Thor moving from 16 ms to 20 ms. On GPU, the release says d1-3B answers a question in under 10 ms and processes a 384px image in under 18 ms on both platforms.

How to use open d1 decision models

These models are aimed at people who need fast, structured decisions, including cases where the input may include images or audio. d1-3B is positioned as the option with the highest decision quality at its size. d1-omni-600M is positioned for cases where footprint matters, meaning its smaller parameter count may fit into tighter deployment budgets.

The setup requires transformers>=5.14. Because the models ship with their own code, they need to be loaded with trust_remote_code=True. For brevity, the release includes only a usage example for d1-3B; instructions for d1-omni-600M are kept in its model card.

Both decision models are open-weight and available on Hugging Face today. The download links are for d1-3B and d1-omni-600M on Hugging Face. Builders can also try the demos in the System One Arcade Hugging Face Space. If someone uses this work, the release asks them to cite the release blog.

Who should use it, and who should wait

For developers building local or edge systems that need quick classification, safety checks, intent routing, visual question answering, or audio-supported decision tasks, the open d1 decision models are directly relevant. d1-3B is the stronger choice when decision quality and multimodal text-image work matter most at its size. d1-omni-600M is worth watching when a smaller footprint is essential, but its early research status means teams should expect the release to be more experimental.

The release is especially useful for people who care about latency on NVIDIA Jetson hardware, because it gives concrete device-level timings. It is less useful for anyone waiting for public vision and audio benchmark numbers, since those are not reported in this version. For now, the clearest takeaway is that d1-3B is the main model to evaluate: it leads the Decision Index 0.2.1 under 10B, scores 82.9 across the seven public datasets, and answers in milliseconds on edge devices.