Liquid AI releases LFM3, six 2B-class open-weight models for devices

The first release includes a 1.7B text model that Liquid says fits in about 1 GB at 4-bit precision and supports a 256K context. The set also has a vision-language model that accepts video input, an audio model for speech in and out across 18 languages, an omni model covering vision, audio and language, and two decision models.

Liquid says the architecture was shaped by a hardware-in-the-loop search across 14 device and backend combinations, and that on-device decoding runs 1.8 to 2.8 times faster than the next-fastest compared model, llama.cpp with Q4_0 at 32K context. Every model can be fine-tuned on a single GPU, and the weights are on Hugging Face. The models are listed as running on llama.cpp, MLX, vLLM and Ollama, among others.

According to Liquid, its vision model averages 71.4 across 62 vision benchmarks, within 2 points of GPT-5.6 Luna at 73.4. The audio model reaches a time to first token as low as 50 ms.

Developers building local agents can run a small model on phones, laptops or edge hardware instead of sending every request to a cloud model. The 1.7B text model has been post-trained for Hermes Agent, Pi and OpenClaw, which may reduce the adaptation work for those setups.

The speed and benchmark figures are Liquid's own numbers, so test them on the hardware and tasks you plan to use before relying on them.