Liquid AI ships LFM3: six 2B-class models on one backbone

Liquid AI published LFM3, its next generation of Liquid Foundation Models, on the Hugging Face blog on October 27. Rather than a single flagship, the release is a restructuring of the lineup: six 2B-class models for text, vision, audio, omni and decision-making, all branched from one LFM3-1.7B-Base backbone — what the company calls “one backbone for six models.”

The blog lists four changes from the LFM2 family: more capability per parameter, every modality on one backbone, an explicit edge-first design, and specialization as the default. Each model fine-tunes on a single GPU; the text model fits in about 1GB at 4-bit precision and carries a 256K context window, which Liquid AI frames as enough to run agents on-device.

The 24-layer base uses a tri-hybrid layer mix — 13 gated short convolutions with a kernel size of 7, five sliding-window attention layers with a 256-token window, and six global attention layers — plus head-wise gated attention that lets a head switch itself off, partial RoPE applied only to the high-frequency half of each head in the global layers, and sandwich normalization on every block. The recipe came from a hardware-in-the-loop architecture search spanning 14 device and backend combinations, from phones and a Raspberry Pi 5 to laptops, cloud CPUs and GPUs, and Qualcomm's Hexagon NPU.

The base was pre-trained on 22T tokens with the same 128K vocabulary as LFM2.5-8B-A1B. The final 3.5T tokens were a mid-training stage weighted toward agentic traces and tool calls, which also stretched the context window to 256K. Liquid AI reports that LFM3-1.7B matches LFM2-1.2B per parameter on short prompts across edge devices and runs about 25% faster with 8K tokens of context.

The six models map onto input types: LFM3-1.7B takes and returns text; LFM3-VL-2B accepts text, images and video; LFM3-Audio-2B takes text and audio and can emit speech; LFM3-Omni-2B covers all four input modalities. The two d1 models sit on top of the vision and omni variants and return what the company calls calibrated answers.

Both new encoders are in-house. The 164M-parameter vision encoder combines convolution and attention and matches Qwen3.5-2B's vision encoder in accuracy with 2.2x fewer FLOPs, while using 3x less compute than the SigLIP2 encoder in earlier LFM vision models; for video it packs each pair of sampled frames into one temporal slice, halving the sequence length. The fully causal audio encoder can prefill while a user is still speaking, with time to first token as low as 50ms, and the new LiquidCodec tokenizer produces speech at roughly 87x real time on a single CPU thread of an Apple M4 Max.

The benchmark claims are the most concrete part of the release. On WebArena-Verified, LFM3-1.7B scores 38.42 against 3.02 for LFM2.5-2.6B, and Liquid AI says it beats Qwen3.5-9B on long-horizon tasks in browsers and business apps. On the vision side, LFM3-VL-2B lifts the average across 62 benchmarks from 57.6 to 71.4 at 1.9B parameters, a comparison the company draws against Gemma 4 12B, which is more than six times larger.

Taken together, the numbers describe a bet against the “bigger and more general” playbook: one small backbone for many modalities, a model chosen by input type, fine-tuning on the user's own data, and deployment wherever the application runs. The models are also post-trained to run inside agent harnesses such as Hermes Agent, Pi and OpenClaw. What happens next will be decided by on-device latency in the wild, how much accuracy survives fine-tuning on narrow data, and whether a six-model family can hold an open-source ecosystem together.