Liquid AI packs six modalities into 2B models: a 4-bit text model near 1 GB, fine-tuned on one GPU

Liquid AI folded text, vision, audio, omni and decision capabilities into one family of 2B-class models, each branching from LFM3-1.7B-Base, so a user picks the model matching their inputs, fine-tunes it on their own data and deploys it where the application runs. The earlier LFM2 family already ran fast on-device with a hybrid of short convolutions and attention, but its modality models were not tied to a single shared base in this way. LFM3-1.7B outperforms LFM2.5-2.6B while using 35% fewer parameters, scoring 38.42 against 3.02 on WebArena-Verified, with the largest gains on agentic tasks.

The backbone was redesigned around agentic use cases, which need more capability per parameter and fast inference over long contexts. LFM2 had shown that a hybrid of short convolutions and attention can run fast on-device without sacrificing quality, and LFM3 rebuilds the layer mix on that basis. LFM3-1.7B-Base has 24 layers: 13 gated short convolutions, 5 sliding-window attention layers with a 256-token window and 6 global attention layers, with convolution kernel size raised from 3 in LFM2 to 7.

The architecture came out of a hardware-in-the-loop search spanning 14 device and backend combinations, from phones and a Raspberry Pi 5 to laptops, cloud CPUs, GPUs and the Qualcomm Hexagon NPU. Earlier architecture search did not cover this breadth of devices and backends. Per parameter, LFM3-1.7B matches LFM2-1.2B on short prompts across edge devices and is about 25% faster with 8K tokens in context.

The vision and audio branches each carry a new in-house encoder, and speech output runs through LiquidCodec, a new audio tokenizer. Previous LFM vision models used the SigLIP2 encoder, and the audio side did not have this in-house encoder and codec pairing. The 164M-parameter vision encoder matches Qwen3.5-2B's vision encoder in accuracy with 2.2x fewer FLOPs under the same LFM2.5-230M language backbone, and LiquidCodec turns generated tokens into speech at about 87x real time on a single CPU thread on Apple M4 Max.

The results cover 2B-class open-weight models in on-device and single-GPU fine-tuning settings, with comparisons against LFM2.5-2.6B, LFM2.5-VL-3B, LFM2.5-Audio-1.5B and d1-3B, plus external models such as Qwen3.5-2B's vision encoder and Qwen2.5-Omni-7B.

The release says d1-omni-2B keeps nearly all of its accuracy when asked by voice instead of in writing but gives no figure, and the 25% speed advantage of LFM3-1.7B over LFM2-1.2B is stated for the 8K-token context condition.