Liquid AI releases open d1-3B decision model: 16 ms on Jetson AGX Thor, 48.57 on Decision Index
Liquid AI released two open-weight decision models, d1-3B (text and image) and the experimental d1-omni-600M (text with image or audio). Unlike generative models, they answer in a single forward pass, and Liquid AI claims d1-3B scores 48.57 on its Decision Index 0.2.1, ahead of all 4B and 9B models it tested.
Liquid AI released two open-weight decision models on October 7, 2026: d1-3B, which takes text and images, and d1-omni-600M, an experimental model that takes text plus either an image or audio. Liquid AI claims d1-3B scores 48.57 on its own Decision Index 0.2.1, the best result under 10B parameters, and that it answers a single question in 16 ms on an NVIDIA Jetson AGX Thor.
What a decision model is
Decision models do not generate tokens. They answer structured questions (yes/no, multiple choice, scored criteria) in a single forward pass. Both models are built on Liquid AI's Liquid Foundation Models (LFMs), but from different backbones:
- d1-3B is trained from LFM2.5-VL-3B, a decoder-only vision-language model. Inputs: text and images.
- d1-omni-600M is trained from LFM2.5-Encoder-350M, a bidirectional encoder, with added vision and audio encoders. Inputs: text and image, or text and audio. Liquid AI labels it an early research release.
Benchmarks (company-reported)
On the Decision Index 0.2.1, Liquid AI says d1-3B scores 48.57, ahead of every 4B and 9B model it tested and of Decider 35B-A3B (47.11). The company ran the comparisons itself, and the index is its own benchmark.
On seven public datasets, d1-3B has a mean of 82.9 and d1-omni-600M has 78.4. The comparison models are Decider 2B (77.1) and Decider 4B (81.1).
| Benchmark | d1-omni-600M | d1-3B | Decider 2B | Decider 4B |
|---|---|---|---|---|
| SQuAD 2.0 | 74.0 | 83.3 | 67.7 | 76.0 |
| 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 |
| PAWS-X | 79.5 | 76.4 | 59.5 | 69.8 |
| Mean | 78.4 | 82.9 | 77.1 | 81.1 |
d1-3B does not win every task. Decider 4B scores higher on MASSIVE intent, BoolQ and XNLI, and d1-omni-600M beats d1-3B on Civil Comments. Liquid AI says d1-omni-600M surpasses Decider 2B with a quarter of the parameters.
Liquid AI reports no vision or audio benchmark scores. It says it validated that d1-3B retains the vision capabilities of its backbone, and that audio decision benchmarks remain an open problem.
Latency
Liquid AI measured d1-3B with NVIDIA on several devices. Speed figures are single-question latency, and no speed numbers were published for d1-omni-600M.
| Device | One question | 384px image | 64 states, packed |
|---|---|---|---|
| Jetson AGX Thor | 16 ms | 35 ms | 262/s |
| Jetson AGX Orin 64 GB | 26 ms | 83 ms | 110/s |
| Jetson Orin Nano | 50 ms | 202 ms | 38/s |
| Apple M5 Pro | 30 ms | 62 ms | 78/s |
| NVIDIA RTX 4090 | 8 ms | 17 ms | 475/s |
| AMD MI325X | 9 ms | 18 ms | 1,106/s |
Processing a 3.4K-token state takes 220 ms on AGX Thor and 1,640 ms on Orin Nano. Three questions over one state cost about 1.3x the time of one on edge devices.
Availability
Both models are on Hugging Face as open weights, with a demo in the System One Arcade Space. They require transformers>=5.14 and trust_remote_code=True. The API exposes named questions of type yes/no, choice or score through system_one and system_one_batch. Context window, training cutoff and license terms were not stated in the announcement.
What this means
The design targets classification, routing and triage workloads, where a generative model's token-by-token decoding adds latency without adding value. Sub-50 ms answers on a Jetson Orin Nano make on-device decisions plausible without a cloud call.
The headline claims need outside validation. The Decision Index is Liquid AI's own benchmark, and the vision split of the newer v0.3 is private. Decider 4B still beats d1-3B on three of seven public tasks, so the advantage is not uniform. The absence of vision and audio scores also leaves multimodal quality unquantified. Teams should test on their own decision tasks before replacing existing classifiers.
Related Articles
Tencent releases Hy4 flagship weights: 770B-parameter MoE, 1,048,576-token context, conditional license
Tencent released downloadable weights for its flagship Hy4 on November 20, with a System Card. The multimodal MoE has a 770-billion-parameter language backbone, 49 billion active parameters per text token and a 1,048,576-token context window, under a license with commercial conditions for model-as-a-service operators.
Tencent releases Hy4 flash: 220B MoE, 12B active parameters, 1M-token context
Tencent has released Hy4 flash, a multimodal Mixture-of-Experts model with 220B total parameters, 12B active per token, and a 1,048,576-token context window. Tencent calls it its most cost-effective Hy model to date and claims a per-task cost of 1/20 that of Hy4 on its internal WorkIndex benchmark.
Baseten releases open-weight bison-1 and groundhog-1 decision models; bison-1 scores 76 on AA Decision Index v2.1
Baseten released bison-1 and groundhog-1, two open-weight decision models built on Gemma 4. Bison-1 scores 76 on the AA Decision Index v2.1, and both models support 128k context and image input.
Thinking Machines Lab releases Inkling 2 Mini: 26B-parameter looped MoE with 3B active and 1M-token context
Thinking Machines Lab has released Inkling 2 Mini, a multimodal Mixture-of-Experts model with about 26B physical total parameters and 3B active per token. It repeats its middle layers three times, supports a 1M-token context, and ships with a DSpark drafter. Hugging Face reports it leads 32 of 48 benchmarks against four comparable models.
Comments
Loading...