Overview
Liquid AI released LFM3 on October 27, 2026, a new generation of open-weight Liquid Foundation Models. The first batch covers six models, all in the 2 billion parameter class, and every one branches from the same backbone, LFM3-1.7B-Base. Text, vision, speech, omni and decision capabilities therefore share one architecture and one fine-tuning workflow.
The backbone is tri-hybrid: of 24 layers, 13 are gated short convolutions, five use sliding-window attention over 256 tokens, and six use full attention. Convolution kernel size rises from 3 in LFM2 to 7, and the design adds head-wise gated attention, a partial RoPE that rotates only the high-frequency half of each head in global layers, and a sandwich norm around each block. Liquid AI arrived at this through hardware-in-the-loop architecture search expanded to 14 device and backend combinations, including Samsung Galaxy S26 Ultra, OnePlus 15, Raspberry Pi 5, Apple M4 Max, AMD Ryzen AI, Arm and x86 cloud CPUs, GPUs and Qualcomm's Hexagon NPU.
Training load is substantial for the size. The backbone saw 22 trillion tokens of pre-training, with the last 3.5 trillion forming a mid-training stage weighted toward agentic traces and tool calls that also stretched context to 256K tokens. Post-training ran in four stages: supervised fine-tuning on 1.75 trillion tokens, mixed reinforcement learning, multi-domain on-policy distillation from six specialist RL teachers, and agentic RL inside real agent harnesses.
Key Features
- One backbone, six models: LFM3-1.7B handles text; LFM3-VL-2B takes text, images and video; LFM3-Audio-2B takes text or speech and answers in text or speech; LFM3-Omni-2B accepts all four modalities and replies in text and speech. Two decision models, d1-2B and d1-omni-2B, return calibrated answers instead of generated text.
- Architecture tuned on hardware: A 14-combination search spanning phones, a Raspberry Pi, laptops, cloud CPUs, GPUs and an NPU shaped the design. Head-wise gating suppresses activation outliers that hurt low-bit quantization, and partial RoPE keeps global attention cheap.
- 256K context, 16 languages: The text model has 1.73B parameters, a 128,000 token vocabulary, a 256K token context and support for 16 languages. At 4-bit precision it occupies about 1GB of memory and runs roughly 25 percent faster per parameter than LFM2-1.2B with 8K tokens in context.
- Smaller model, higher scores: Liquid AI reports LFM3-1.7B beating its LFM2.5-2.6B predecessor with 35 percent fewer parameters: 38.42 versus 3.02 on WebArena-Verified. The vision model lifts the average across 62 benchmarks from 57.6 to 71.4 while dropping from 3.1B to 1.9B parameters.
- Decision models with probabilities: d1-2B and d1-omni-2B answer a set of questions in a single forward pass and return calibrated probabilities instead of free-form text, which suits ticket triage or image QA where a threshold decides the route.
- Fine-tune on a single GPU: Every model can be fine-tuned on one GPU, and Liquid AI publishes notebooks for text, vision, audio and omni covering SFT, DPO, GRPO and continued pre-training.
Use Cases
- Private agents on a phone or laptop, using the 1.7B text model at about 1GB in 4-bit.
- Receipt, invoice and screenshot reading, turning images straight into structured JSON with the VL model.
- Spoken commands mapped to function calls, with recognition, synthesis and translation across 18 languages in the audio model.
- Meeting recordings that need a single multimodal pass, using the omni model for questions and spoken replies.
- Threshold-driven routing such as ticket classification or content review, using the decision models' calibrated probabilities.
- Raspberry Pi, NPU and edge boxes, all of which are in the hardware matrix Liquid AI searched over.
Pros
- Six models share one backbone, so tooling and fine-tuning recipes carry over.
- The 1.7B text model fits in about 1GB at 4-bit, workable on an 8GB device.
- A 256K token context is unusual at this parameter count.
- Liquid AI reports beating the prior generation with 35 percent fewer parameters and a nearly 14 point lift on vision averages.
- Decision models ship calibrated probabilities, removing a calibration step from the pipeline.
- Open weights on Hugging Face plus a playground and API, with every model fine-tunable on one GPU.
Pricing
Weights are released under the LFM Open License v1.0 and available on Hugging Face, with hosted access through the Liquid AI Playground and API. The licence covers commercial use for organisations below $10 million in annual revenue; at or above that threshold a commercial licence is required. Self-hosting carries only hardware cost, and Liquid AI says the 4-bit text model needs about 1GB of memory. API pricing and free tiers follow the Liquid AI site.
Summary
LFM3 splits general capability into six small specialised models sharing one backbone. For on-device work the value is optionality rather than benchmark scores: one fine-tuning workflow covers text, vision, speech and decisions, model sizes sit between 1GB and 2GB, and phones, Raspberry Pi and NPUs are all in the validated backend list. Liquid AI is also clear about the ceiling, since a 2B class model is not where long multi-step reasoning or complex coding belongs.
Version History
- LFM3 ships six open-weight 2B-class models on one shared backbone (2026-10-27): Liquid AI released the first six LFM3 open-weight models, all branched from LFM3-1.7B-Base: text, VL, audio, omni and the decision models d1-2B and d1-omni-2B. The backbone is tri-hybrid with 13 gated short convolutions, five 256-token sliding-window attention layers and six full-attention layers across 24 layers. Trained on 22 trillion tokens, the last 3.5 trillion on agentic data that extended context to 256K. The text model has 1.73B parameters, 16 languages, about 1GB at 4-bit, and scores 38.42 on WebArena-Verified while using 35 percent fewer parameters than LFM2.5-2.6B. Released under the LFM Open License v1.0.