What changes when six AI models share one 2-billion-parameter backbone?

Liquid AI calls LFM3 a "starting point for specialized intelligence." The phrase means you download one compact backbone, pick the modality you need, and adapt it to your data without leaving a single GPU.

The backbone is LFM3-1.7B-Base: 24 layers, 1.7 billion parameters, 256K context window, 22 trillion pre-training tokens. Six models branch from it. LFM3-1.7B handles text and tool calling. LFM3-VL-2B adds images and video. LFM3-Audio-2B adds speech input and output. LFM3-Omni-2B combines vision, audio, and text in both directions. d1-2B and d1-omni-2B replace generated answers with calibrated probabilities for decision tasks.

Every model fits in the 2B parameter neighborhood. The text model quantizes to roughly 1 GB at 4-bit precision. Fine-tuning notebooks ship for each variant. The company says you can specialize any of them on your own terminology, tools, and deployment target in a single training run.

Tri-hybrid layer mix and why it matters

LFM2 already mixed short convolutions with attention. LFM3 redesigns that mix through what the team calls hardware-in-the-loop architecture search across 14 device and backend combinations — phones, Raspberry Pi 5, laptops, cloud CPUs, GPUs, and the Qualcomm Hexagon NPU.

The resulting 24-layer stack contains 13 gated short convolution layers with kernel size 7, five sliding-window attention layers with a 256-token window, and six global attention layers. Each attention head gets a sigmoid gate that can shut the head off when it contributes nothing, a feature the authors say helps suppress activation outliers that hurt quantization. Global attention layers rotate only the highest-frequency half of each head via partial RoPE; sliding-window layers keep full RoPE. Every block normalizes both input and output — sandwich norm — to keep activations well-scaled.

At 8K tokens in context, LFM3-1.7B runs about 25 percent faster per parameter than its predecessor on edge devices. The final 3.5 trillion tokens of pre-training used a curated mix heavier on agentic traces and tool calls, which also extended the context window from 128K to 256K tokens.

New encoders cut compute for vision and audio

The vision encoder combines convolution and attention in a 164M-parameter module. Liquid AI reports it matches the accuracy of Qwen3.5-2B's vision encoder while requiring 2.2x fewer FLOPs when both are measured with the same LFM2.5-230M language backbone. For video, each pair of sampled frames embeds into one temporal slice, halving the temporal sequence length.

The audio encoder mixes sliding-window attention and short convolutions in a fully causal design. That means the model can prefill while the user speaks, reaching time-to-first-token as low as 50 ms. LiquidCodec, the new audio tokenizer, turns generated tokens into speech at about 87x real time on a single CPU thread on Apple M4 Max and supports streaming decoding.

Decision models output probabilities, not tokens

d1-2B and d1-omni-2B answer structured questions over a single input — text, images, video, or audio — in one forward pass. Question types include yes/no with calibrated probability, single choice, multiple choice, scored urgency, and now object pointing, value extraction, numeric ranges, and multi-select. d1-2B replaces the earlier d1-3B with a third fewer parameters and higher scores on all four internal decision suites; for example, 75.5 versus 64.0 on ImajevBench. d1-omni-2B adds audio input and retains nearly all accuracy when questions are spoken instead of written.

Fine-tuning on one GPU is the default path

The release includes Colab notebooks for supervised fine-tuning, DPO, GRPO, and continued pre-training across the family. Tasks range from extracting JSON from messy documents and transcribing equations to LaTeX, to mapping spoken smart-home commands to function calls and teaching the model a new voice. A robotics demo jointly trains three complementary models — orchestrator, navigator, and manipulation VLA — from LFM3-VL-2B across 15 embodiments, achieving 96.2 percent average success on LIBERO suites and 45.7 percent on RoboCasa-GR1 tabletop tasks with action chunks generated in under 50 ms on a Jetson AGX Orin with TensorRT FP16.

All weights are open and hosted on Hugging Face. The collection includes LFM3-1.7B-Base for custom post-training, plus guides for running local agents such as OpenClaw, Hermes Agent, and Pi.