Liquid AI's d1-2B trails d1-3B on 3 Japanese benchmarks

Liquid AI added d1-2B to its open-weight d1 family, with the Hugging Face repository created on October 24 and the first commit on October 27 — roughly three weeks after d1-3B and d1-omni-600M appeared on October 5. d1-2B is a decision model rather than a text generator: it returns a probability-scored answer in one pass, taking state (text or JSON), images and video, and questions such as yes/no, multiple choice, graded rating, value extraction and object location. It has 1.92B parameters, was post-trained from Liquid AI's LFM3-VL-2B vision-language model, supports 16 languages including Japanese, and is under the LFM Open License v1.0; the model card does not state a release date.

The comparison is not a like-for-like new measurement of every model. Only d1-2B was measured for this article, while d1-3B and d1 API figures come from earlier runs, and the author notes the possibility that the API or weights changed in between. The model card's table and this article's results can appear to conflict, but the benchmarks differ in content, so the author does not treat that as a contradiction — limited to Japanese sentence relations, commonsense and article classification, d1-3B was more accurate. Against other models, d1-2B sits close to Jeff Gemma4-E2B among locally run weights; the author did not run significance tests on those three differences. ModernBERT-Ja-310M was fine-tuned on each benchmark's training data, making its conditions different from the others.