Tool Suggestions
Liquid AI LFM3: Six Open-Weight 2B Models That Run On-Device, Including a 1GB Phone Agent
Liquid AI released LFM3 on October 27, a family of six open-weight models of about 2 billion parameters each, all built on one shared base and all designed to run on local hardware instead of in a data centre. The headline model is the text model, LFM3-1.7B: 1.73 billion parameters, a 256,000-token context window, and a footprint of about 1GB at 4-bit precision — small enough for a phone or a laptop without a dedicated GPU.
The six models
All six branch from the same foundation, LFM3-1.7B-Base, pre-trained on 22 trillion tokens. The lineup is: LFM3-1.7B for text, built for local agents; LFM3-VL-2B for vision and video; LFM3-Audio-2B for voice interaction in 18 languages; LFM3-Omni-2B, which handles text, images, video and audio together; and two decision models, d1-2B and d1-Omni-2B, which give calibrated single-pass outputs for fast, non-generative decisions. Rather than running one large general-purpose model, a developer picks the model that fits the job, customises it and runs it on-device.
Why the architecture matters
The shared backbone is what Liquid calls tri-hybrid: 24 layers split between 13 gated short convolutions, 5 sliding-window attention layers and 6 global attention layers, found through a hardware-in-the-loop search across 14 device and backend combinations, including a Raspberry Pi 5 and a Qualcomm Hexagon NPU. That is a practical difference, not a theoretical one: the architecture was tuned against the actual chips a phone or a laptop ships with, rather than against a GPU cluster most developers cannot afford to rent. Liquid says the text model is about 25% faster than its predecessor LFM2 at 8,000 tokens of context, and that it is post-trained to run in agent harnesses for jobs like inbox triage, calendars and web research.
What it costs and where to get it
Base and post-trained weights are available as open-weight models on Hugging Face and on Liquid AI’s platform, with GGUF, MLX, ONNX, FP8 and NVFP4 checkpoint variants for llama.cpp, MLX, vLLM, SGLang and ONNX Runtime — llama.cpp support was already merged at release. Every model can be fine-tuned on a single GPU, which is the point: a team can adapt a model to its own data without renting a cluster. Under the LFM Open License v1.0, commercial use is free below $10 million in annual revenue; companies at or above that threshold need a separate commercial agreement with Liquid AI.
What it can’t do
Liquid is unusually candid about the limits, and they matter. Its own model card says the small text model is not intended for agentic coding or complex multi-step reasoning — its strength is local agents, tool calling, structured outputs and domain fine-tuning. Benchmark figures showing the small models competing with models several times their size are Liquid’s own and have not been independently verified, and larger models still win on knowledge and on coding, where size matters most. The honest pitch is narrower and better: a model small enough to live on your device is now good enough for a large share of daily work, and since on-device tokens cost only electricity, dozens of agents can run continuously in the background.