Liquid AI Releases Six Compact LFM3 Models for On-Device Enterprise AI

Quick Facts

• Liquid AI released six open-weight models under the LFM3 family, all built on a shared 1.7-billion-parameter backbone with specialized versions reaching about 2 billion parameters.
• The models support text, vision, video, audio, and structured decision-making, and are designed to run on devices as small as a smartphone or Raspberry Pi 5.
• LFM3-1.7B scores 38.42 on WebArena-Verified, compared with 28.94 for Qwen3.5-9B and 3.02 for Liquid's prior LFM2.5-2.6B model.

Liquid AI launched the LFM3 model family on Oct. 27, giving enterprise teams six compact, open-weight models designed to run browser agents, process voice commands, and recognize images on company-controlled hardware without relying on external cloud services.

The six models share a common backbone and cover four modalities: text, vision with video, audio, and a combined omni format. A separate pair of decision models handles structured classification tasks.

Architecture

LFM3 runs on what Liquid calls a tri-hybrid architecture, combining gated short convolutions, sliding-window attention, and global attention across 24 layers. The base model trained on 22 trillion tokens, including 3.5 trillion mid-training tokens focused on agentic traces and tool calls, extending the context window to 256K tokens.

Post-training ran in four stages: supervised fine-tuning on 1.75 trillion tokens, mixed reinforcement learning, multi-domain distillation from six specialist teachers, and agentic reinforcement learning inside real agent harnesses. Liquid used a hardware-in-the-loop architecture search covering 14 device and backend combinations to optimize performance across phones, laptops, cloud CPUs, GPUs, and the Qualcomm Hexagon NPU.

The Six Models

LFM3-1.7B handles text in and text out. At 4-bit precision, it requires about 1 GB of memory, fitting on any phone or laptop. LFM3-VL-2B adds vision and video input and is designed for on-device visual tasks, including object detection and spatial reasoning suited for robotics applications.

LFM3-Audio-2B accepts text or audio and responds with text, speech, or both across 18 languages. LFM3-Omni-2B combines all modalities, taking text, images, video, and audio as input and replying in text or speech.

The d1-2B and d1-omni-2B models are built for structured decisions rather than generated text. They take a situation, such as a support ticket, invoice, app screen, or voice note, and return calibrated probability scores across a set of questions in a single forward pass. They can flag objects, read values from documents, estimate numbers with a range, and select multiple applicable options at once.

Performance

On vision benchmarks, LFM3-VL-2B raises the average across 62 vision benchmarks from 57.6 to 71.4 compared with LFM2.5-VL-3B, using 1.9 billion parameters against 3.1 billion. Liquid says that result puts it on par with Gemma 4 12B, a model more than six times larger.

LFM3-1.7B beats the prior LFM2.5-2.6B while using 35% fewer parameters, posting 45.60 on LiveClaw, 71.39 on BFCL V4 tool-calling, and 96.66 on IFEval. The d1-2B model replaces d1-3B with a third fewer parameters and scores 75.5 against 64.0 on ImajevBench.

On-device speed tests on an AMD Ryzen AI Max+ 395, an Apple M5 Max, and an iPhone 17 Pro show generation speeds between 1.8 and 2.8 times those of the next-fastest model in each comparison, using llama.cpp at 4-bit quantization. On a single NVIDIA B300 with SGLang, LFM3-1.7B produces 3.6 billion tokens per day.

Early benchmarks on Arm CPUs show a 3.1x speedup in prefill and 1.5x speedup in decode with SME2 enabled, reaching over 400 tokens per second in prefill using two CPU threads.

Business Case

Liquid AI CEO Ramin Hasani said the company views on-device AI as a distinct market. "We need this kind of intelligence on the edge as well as cloud AI to truly do a planet-scale deployment of AI for good," he said.

The models are open-weight and available on Hugging Face, with fine-tuning possible on a single GPU. That combination of small footprint, multimodal support, and open weights positions LFM3 for enterprise teams that need AI capabilities without sending sensitive data to external providers.