d1 Decision Models
modelYour notes
First-in-series of a model class Liquid calls a decision model, and the distinction is architectural rather than positioning: these "don't produce tokens. Instead, they produce an answer in a single forward pass." You pass a state — text, JSON, an image, audio — plus named questions, and get back calibrated typed answers with zero output tokens. That makes latency a property of the input alone, which is the point: 8 ms per question on an RTX 4090, 26 ms on Jetson AGX Orin, 50 ms on Orin Nano. It is explicitly not a chat model and cannot be used as one.
Two open-weight variants, both post-trained from Liquid's own LFM2.5 family rather than trained from scratch. d1-3B (3.12B) comes from LFM2.5-VL-3B, a decoder-only VLM, by weight averaging and fine-tuning the text and vision backbones; it takes text and images. d1-omni-600M comes from LFM2.5-Encoder-350M — a bidirectional encoder, which is the more interesting choice, since a decision model never decodes — with a FastConformer audio encoder and adapter, then vision via an adapter plus LoRA. It takes text with either an image or audio.
On the seven text benchmarks Liquid reports, d1-3B means 82.9 against Decider 4B's 81.1, with SQuAD 2.0 at 85.3; d1-omni-600M reaches 95.8 on Civil Comments toxicity. Read the headline "Decision Index" figures with care (d1-3B 48.57, d1-omni-600M 15.95 on the v0.2.1 public split): that index is a third-party user repository (github.com/apolinario/decision-index), not a Liquid benchmark and not an established one, so it should not be read as an independent evaluation of the class Liquid is simultaneously defining. The arxiv:2511.23404 tag on the model cards is the LFM2 tech report, not a d1 paper — there is no d1 paper at filing.
The second generation arrived with LFM3 on October 24–27, 2026. d1-2B is built on LFM3-VL-2B and takes text, images and video; d1-Omni-2B is built on LFM3-Omni-2B and adds audio, with spoken questions scoring within 0.5 points of written ones on unseen tasks, per its card. Liquid's card says d1-2B leads the open d1 models on its board of over a hundred decision benchmarks, ahead of d1-3B at under two-thirds the size, and quotes 5.7 ms per decision on an H100. Beyond classification, both also extract spans, locate objects and estimate quantities with an interval.
Model Details
Benchmark Scores
| Benchmark | Score | Mode |
|---|---|---|
| Seven-benchmark text mean (d1-3B) | 82.9 | — |
| SQuAD 2.0 (d1-3B) | 85.3 | — |
| Civil Comments toxicity (d1-omni-600M) | 95.8 | — |
Variants
| Name | Parameters | Notes |
|---|---|---|
| d1-3B | 3.12B | from LFM2.5-VL-3B; text + image input |
| d1-omni-600M | 600M | from LFM2.5-Encoder-350M, a bidirectional encoder; text + image or audio |
| d1-2B | — | 2026-10-24; from LFM3-VL-2B; text, images and video; 5.7 ms per decision on H100 |
| d1-Omni-2B | — | 2026-10-24; from LFM3-Omni-2B; text, images and audio |