Back to the ticker

Liquid AI releases d1-3B and d1-omni-600M as open-weight decision models

Liquid AI has released two open-weight decision models, d1-3B and d1-omni-600M, which answer a question without generating any tokens. Both produce their answer in a single forward pass, so a yes/no, multiple-choice or scale question costs one pass through the network instead of a decoded response. The company says d1-3B scores 48.57 on the public split of the Decision Index v0.2.1, ahead of every model under 10B parameters and on par with Decider 35B-A3B, a decision model 12 times its size. Liquid AI had previously offered d1 through its API only.

A single question takes 16 ms on an NVIDIA Jetson AGX Thor, 26 ms on a Jetson AGX Orin 64 GB, 30 ms on an Apple M5 Pro and 50 ms on a Jetson Orin Nano, measured end to end from input to output. Three questions cost about 1.3 times as much as one, going from 16 ms to 20 ms on the AGX Thor, and a 384px image adds 35 ms there. On desktop hardware the same model answers in 8 ms on an RTX 4090 and 9 ms on an AMD MI325X. Liquid AI ran the tests one request at a time and reports latency for d1-3B only.

Seven public text benchmarks cover reading comprehension, toxicity detection, intent classification, medical question answering and cross-lingual understanding. d1-3B reaches a mean of 82.9 across them, ahead of Decider 4B at 81.1, while d1-omni-600M lands at 78.4 and beats Decider 2B at a quarter of the parameters. Liquid AI reports no vision scores because the vision split of the Decision Index v0.3 is private, and it calls dedicated audio decision benchmarks an open problem.

Text benchmarks, d1-3B against Decider 4B

Benchmarkd1-3BDecider 4B
SQuAD 2.085.376.0
Civil Comments93.092.8
MASSIVE intent87.388.3
PubMedQA66.063.3
BoolQ86.789.0
XNLI85.088.6
PAWS-X76.969.8
Mean82.981.1

All figures according to Liquid AI. d1-omni-600M reaches a mean of 78.4 on the same seven benchmarks.

Both models start from existing LFM2.5 checkpoints. d1-3B is an average of LFM2.5-2.6B and the text backbone of LFM2.5-VL-3B, fine-tuned with different random seeds and data mixtures and then merged again, and it takes text and images as input. d1-omni-600M builds on the bidirectional LFM2.5-Encoder-350M and adds a FastConformer audio encoder and the vision encoder of LFM2.5-VL-450M in stages, so it accepts text with either an image or audio. Liquid AI calls it a first experimental checkpoint that is under active development.

Architecture diagram of d1-3B in which yes/no, choice and score prompts go through a tokenizer and an image through a 27-layer SigLIP2 vision encoder into a 30-layer causal LFM2.5-2.6B trunk, whose LM head outputs probabilities per answer option

The question types are encoded alongside the image, and the head returns a probability for every answer option.Liquid AI

Day-one support for llama.cpp covers Apple, Qualcomm, AMD and NVIDIA hardware, with NVFP4 on the NVIDIA stack from DGX in the data center down to Jetson at the edge, and Liquid AI documents the local setup. The weights can be downloaded, fine-tuned and deployed without restrictions according to the company, which names no license in the announcement. Ten browser demos run d1-3B in a loop over live camera input in a Hugging Face space, among them gesture-controlled games and live content moderation, and NVIDIA lists both models in its Jetson AI Lab. A further demo has the model steering a simulated robot arm in NVIDIA Isaac Sim, served from a Jetson.

A simulated robot arm sorts coloured blocks into matching trays in NVIDIA Isaac Sim, played back at 4x speed. Open video
  1. Liquid AI releases a DSpark drafter that speeds up LFM2.5-VL-3B decoding
  2. Liquid AI releases LFM2, three CPU-first models from 350M to 1.2B
  3. Hugging Face's SmolLM3 is a 3B model with 128k context and two reasoning modes