$ cat wiki/models/d1-3b.md
d1-3B
Spec
| Attribute | Value |
|---|---|
| Developer | Liquid AI |
| Released | 2026-10-07 |
| Announced | 2026-10-07 |
| Context window | unknown |
| Pricing | unknown |
| License | unknown |
| Availability | Downloadable weights on Hugging Face; llama.cpp support |
| The announcement establishes weight availability, but does not identify a license or context limit. Those fields remain unknown pending a model-specific source. (source) |
Release Date
Released on 2026-10-07 as an open-weight decision model accepting text and images. (source)
Benchmarks
Liquid AI reports 48.57 on the Decision Index v0.2.1 public split. It does not report the private vision split from v0.3 or a dedicated audio decision benchmark. (source)
The vendor measures one question at a time at 8 ms on NVIDIA RTX 4090 and 16 ms on Jetson AGX Thor. On the latter, a 3.4K-token state takes 220 ms and a 384px image takes 35 ms. Input shape and hardware are part of each result; these are not interchangeable latency claims. (source)
Use Cases
The model returns a decision in one forward pass without generating output tokens. It is trained from the decoder-only LFM2.5-VL-3B, for structured decisions using multimodal inputs. (source)
Compared To
- d1-omni-600M is the other checkpoint in the same release; the two use different backbones and input combinations. (source)
- Jev supplies a related decision-model comparison in the wiki. The release's public-split score does not establish a matched comparison with every Jev deployment. (source)