ENPT

Liquid AI's LFM3 brings two open 2B decision models: d1-2B and d1-Omni-2B

Last updated:

Source · Model card: d1-3B

publicado (OK do Fabricio 28/10 BRT).

Liquid AI announced LFM3 on 27 Oct 2026 (14:07 BRT), a family of six open models around 2B parameters on a new 24-layer backbone that mixes short convolutions with sliding-window and global attention. Two of them are decision models:

  • d1-2B: 1.92B, built on LFM3-VL-2B, text or JSON plus pictures or video.
  • d1-Omni-2B: 2.17B, built on LFM3-Omni-2B, adds audio.

Both return typed answers with probabilities in one forward pass, with 128K context and 16 languages, under the LFM Open License v1.0 (not Apache). The launch post had about 7.3k likes and 1.4M views within a day.

Vendor numbers (UnverifiedClaim). The d1-2B card says it leads the open d1 models on Liquid’s own board of over a hundred decision benchmarks, ahead of d1-3B, and answers in 5.7 ms on an NVIDIA H100. Neither model is on the Decision Index or JevBench yet. For comparison, d1-3B sits at Full 31.12 on the Decision Index 0.4.1.

The hosted d1 API is a different product; Liquid has not said whether it shares weights with these checkpoints.