Image: Liquid AI

Liquid AI has released LFM3, a new family of open weight models that tries to squeeze agent skills, vision, speech and video into models small enough for a phone. The company announced six models on October 27, all in the 2 billion parameter class and all trained from one shared base called LFM3-1.7B-Base.

Introducing LFM3: our new generation of Liquid Foundation Models. With a new tri-hybrid architecture and new multimodal encoders, the LFM3 family is the ideal model suite for specialized intelligence.

— Liquid AI (@liquidai) October 27, 2026

What is in the family

The lineup covers every input type from one backbone. LFM3-1.7B handles text. LFM3-VL-2B reads images and, for the first time in a Liquid vision model, video. LFM3-Audio-2B listens and speaks. LFM3-Omni-2B, the company's first omni model, takes text, images, video and audio together and can answer out loud. Two decision models, d1-2B and d1-omni-2B, return calibrated answers to a list of questions in a single pass instead of writing free text.

The LFM3 model family. Image: Liquid AI

How it stacks up

The headline claim is that LFM3-1.7B beats its own predecessor, LFM2.5-2.6B, with 35% fewer parameters. The model card compares it with MiniCPM5 2B, Gemma 4 E4B, Qwen3.5 4B and Qwen3.5 9B. Liquid AI reports an average of 49.37 across its full table, against 47.24 for Qwen3.5 9B, a model more than five times larger.

The gap is widest on agent work. On WebArena-Verified the new model scores 38.42, against 3.02 for LFM2.5-2.6B and 28.94 for Qwen3.5 9B. On LiveClaw it posts 45.60 against 35.77 for Qwen3.5 9B, and on the BFCL V4 tool calling test it reaches 71.39 against 62.83.

The picture is less rosy elsewhere. On GPQA Diamond it trails Qwen3.5 9B by 63.37 to 80.40, and on RULER at 128K tokens it scores 65.56 against 87.11. Liquid AI itself says the text model is not recommended for agentic coding or complex multi step reasoning, and it advises pairing it with retrieval for knowledge heavy tasks.

The vision model also makes a strong claim. According to Liquid AI, LFM3-VL-2B lifts its average across 62 vision benchmarks from 57.6 to 71.4, putting it on par with Gemma 4 12B, a model more than six times its size.

LFM3-Omni-2B benchmark coverage. Image: Liquid AI

What hardware it runs on

Liquid AI says the text model fits in about 1 GB at 4 bit precision and has a 256K token context window. The company tuned the design with hardware in the loop search across 14 device and backend combinations, including the Galaxy S26 Ultra, OnePlus 15, Raspberry Pi 5, Apple M4 Max, AMD Ryzen AI chips, cloud CPUs, GPUs and the Qualcomm Hexagon NPU. It claims the model runs about 25% faster than LFM2-1.2B per parameter once 8K tokens sit in context, and that its audio encoder can start answering in as little as 50 ms.

LFM3-1.7B edge inference results. Image: Liquid AI

How to try it and what it costs

All six models are on Hugging Face in the LFM3 collection, with GGUF, ONNX, MLX and FP8 versions, and they run in Transformers, llama.cpp and vLLM. Liquid AI also offers the models through its Playground and API, and it publishes a guide for running them inside local agents such as OpenClaw, Hermes Agent and Pi. Every model can be fine tuned on a single GPU, and the Hugging Face post links ready made notebooks for tasks like receipt reading and phone app tool calling.

Licensing is the main caveat. The models use the LFM Open License v1.0, which is free for research and for commercial use only while a company stays under $10 million in annual revenue. Larger businesses need a separate deal with Liquid AI, and the company did not list API prices in its launch materials at the time of this writing.

For developers building offline assistants, voice features or robots, LFM3 is one of the more capable small model families available, though the benchmark figures come from Liquid AI and have not yet been checked independently.