Liquid AI Launches Open Multimodal d1 Decision Models for Edge AI
Liquid AI releases open d1-3B and experimental d1-omni-600M, bringing very fast text, vision and audio decisions to edge hardware with strong benchmarks.
Summary
Liquid AI released the open-weight d1-3B and experimental d1-omni-600M on October 7, 2026, with downloads and System One Arcade demos on Hugging Face. Unlike generative models, both make structured decisions in one forward pass without producing tokens. d1-3B uses the decoder-only LFM2.5-VL-3B for text and images; d1-omni-600M uses the bidirectional LFM2.5-Encoder-350M plus vision and audio encoders for text with either images or audio.
On Decision Index 0.2.1, d1-3B scored 48.57, beating all tested 4B and 9B models and Decider 35B-A3B at 47.11. Across SQuAD 2.0, Civil Comments, MASSIVE intent, PubMedQA, BoolQ, XNLI and PAWS-X, d1-omni-600M and d1-3B respectively scored 74.0/83.3, 95.8/93.3, 86.1/86.9, 61.3/68.3, 77.7/86.3, 74.7/85.6 and 79.5/76.4. Their means were 78.4 and 82.9, versus Decider 2B's 77.1 and Decider 4B's 81.1; d1-omni used one-quarter of Decider 2B's parameters.
For one question, three questions, a 3.4K-token state, a 384px image and 64 packed states, latency and throughput were 30/41/640/62 ms and 78/s on Apple M5 Pro; 16/20/220/35 ms and 262/s on Jetson AGX Thor; 26/35/560/83 ms and 110/s on Jetson AGX Orin 64 GB; 50/73/1,640/202 ms and 38/s on Jetson Orin Nano; 8/21/102/17 ms and 475/s on NVIDIA RTX 4090; and 9/14/44/18 ms and 1,106/s on AMD MI325X. Liquid AI withheld d1-omni speed results and public vision or audio benchmarks because it remains early research, Decision Index v0.3 has only a private vision split, and audio decision benchmarking remains unresolved. Use requires transformers 5.14 or later and trust_remote_code=True.
Positives
- d1-3B scored 48.57 on Decision Index 0.2.1, ahead of every tested 4B and 9B model and Decider 35B-A3B at 47.11.
- d1-3B achieved the highest seven-dataset mean, 82.9, exceeding Decider 4B's 81.1.
- d1-omni-600M scored 78.4, beating Decider 2B's 77.1 with one-quarter as many parameters.
- Jetson AGX Thor answered one question in 16 ms and handled 64 packed states at 262 per second.
- AMD MI325X processed one question in 9 ms and 64 packed states at 1,106 per second.
- Both open-weight models are downloadable on Hugging Face, with demos available through System One Arcade.
Risks & concerns
- d1-omni-600M remains an early research release, and Liquid AI provides no inference-speed measurements for it.
- Decision Index v0.3 offers only a private vision split, so Liquid AI reports no public vision benchmark results.
- Audio decision benchmarks remain an open problem, leaving d1-omni-600M's audio quality unquantified.
- d1-3B's 86.9, 86.3 and 85.6 on MASSIVE intent, BoolQ and XNLI trail Decider 4B's 88.3, 89.0 and 88.6.
- Jetson Orin Nano needed 1,640 ms for a 3.4K-token state and 202 ms for a 384px image.