Liquid AI released two open-weight decision models on October 7, 2026: d1-3B and experimental d1-omni-600M. The company says d1-3B is the highest-scoring decision model under 10B parameters on Decision Index 0.2.1, with 48.57 points, ahead of every 4B and 9B model and Decider 35B-A3B at 47.11. d1-3B accepts text and images; d1-omni-600M accepts text with images or text with audio.

The models are built on Liquid Foundation Models and make decisions in a single forward pass, rather than generating tokens. d1-3B uses the decoder-only LFM2.5-VL-3B backbone. d1-omni-600M uses the bidirectional LFM2.5-Encoder-350M backbone, with added vision and audio encoders; Liquid AI says it remains an early research release under development.

Across seven public datasets covering reading comprehension, toxicity detection, intent classification, medical question answering and cross-lingual understanding, d1-3B scored a mean 82.9, above Decider 4B's 81.1. d1-omni-600M scored 78.4, exceeding Decider 2B's 77.1 with one-quarter of its parameters. The reported dataset results for d1-3B, Decider 2B, Decider 4B and Decider 35B-A3B, respectively, were: SQuAD 2.0, 83.3, 67.7, 76.0, with no d1-3B score shown in the table; Civil Comments, 95.8, 93.3, 93.6, 92.8; MASSIVE intent, 86.1, 86.9, 81.1, 88.3; PubMedQA, 61.3, 68.3, 65.7, 63.3; BoolQ, 77.7, 86.3, 87.3, 89.0; XNLI, 74.7, 85.6, 85.0, 88.6; and PAWS-X, 79.5, 76.4, 59.5, 69.8. Liquid AI says d1-3B retains its backbone's vision capabilities and d1-omni-600M handles all three modalities, but reports no vision or audio benchmark scores.

With NVIDIA, Liquid AI measured d1-3B on edge devices: one-question response times were 16 ms on Jetson AGX Thor, 26 ms on Jetson AGX Orin 64 GB and 50 ms on Jetson Orin Nano; Apple M5 Pro took 30 ms. Three questions took 20 ms, 35 ms, 73 ms and 41 ms, respectively. For a 3.4K-token state, times were 220 ms, 560 ms, 1,640 ms and 640 ms; processing a 384px image took 35 ms, 83 ms, 202 ms and 62 ms. Throughput for 64 packed states was 262/s, 110/s, 38/s and 78/s, respectively. On NVIDIA RTX 4090 and AMD MI325X, one-question times were 8 ms and 9 ms, three-question times 21 ms and 14 ms, 3.4K-token-state times 102 ms and 44 ms, image times 17 ms and 18 ms, and throughput 475/s and 1,106/s. d1-omni-600M has no speed results in this release.

Liquid AI positions the models for fast structured decisions with multimodal inputs: d1-3B for decision quality at its size and d1-omni-600M for smaller footprint. Both are open-weight and available on Hugging Face; usage requires transformers>=5.14 and loading with trust_remote_code=True. The company also directs users to demos in its System One Arcade Hugging Face Space.

Liquid AI released two open-weight decision models on October 7, 2026: d1-3B and experimental d1-omni-600M. The company says d1-3B is the highest-scoring decision model under 10B parameters on Decision Index 0.2.1, with 48.57 points, ahead of every 4B and 9B model and Decider 35B-A3B at 47.11. d1-3B accepts text and images; d1-omni-600M accepts text with images or text with audio.

The models are built on Liquid Foundation Models and make decisions in a single forward pass, rather than generating tokens. d1-3B uses the decoder-only LFM2.5-VL-3B backbone. d1-omni-600M uses the bidirectional LFM2.5-Encoder-350M backbone, with added vision and audio encoders; Liquid AI says it remains an early research release under development.

Across seven public datasets covering reading comprehension, toxicity detection, intent classification, medical question answering and cross-lingual understanding, d1-3B scored a mean 82.9, above Decider 4B's 81.1. d1-omni-600M scored 78.4, exceeding Decider 2B's 77.1 with one-quarter of its parameters. The reported dataset results for d1-3B, Decider 2B, Decider 4B and Decider 35B-A3B, respectively, were: SQuAD 2.0, 83.3, 67.7, 76.0, with no d1-3B score shown in the table; Civil Comments, 95.8, 93.3, 93.6, 92.8; MASSIVE intent, 86.1, 86.9, 81.1, 88.3; PubMedQA, 61.3, 68.3, 65.7, 63.3; BoolQ, 77.7, 86.3, 87.3, 89.0; XNLI, 74.7, 85.6, 85.0, 88.6; and PAWS-X, 79.5, 76.4, 59.5, 69.8. Liquid AI says d1-3B retains its backbone's vision capabilities and d1-omni-600M handles all three modalities, but reports no vision or audio benchmark scores.

With NVIDIA, Liquid AI measured d1-3B on edge devices: one-question response times were 16 ms on Jetson AGX Thor, 26 ms on Jetson AGX Orin 64 GB and 50 ms on Jetson Orin Nano; Apple M5 Pro took 30 ms. Three questions took 20 ms, 35 ms, 73 ms and 41 ms, respectively. For a 3.4K-token state, times were 220 ms, 560 ms, 1,640 ms and 640 ms; processing a 384px image took 35 ms, 83 ms, 202 ms and 62 ms. Throughput for 64 packed states was 262/s, 110/s, 38/s and 78/s, respectively. On NVIDIA RTX 4090 and AMD MI325X, one-question times were 8 ms and 9 ms, three-question times 21 ms and 14 ms, 3.4K-token-state times 102 ms and 44 ms, image times 17 ms and 18 ms, and throughput 475/s and 1,106/s. d1-omni-600M has no speed results in this release.

Liquid AI positions the models for fast structured decisions with multimodal inputs: d1-3B for decision quality at its size and d1-omni-600M for smaller footprint. Both are open-weight and available on Hugging Face; usage requires transformers>=5.14 and loading with trust_remote_code=True. The company also directs users to demos in its System One Arcade Hugging Face Space.