open-source

Liquid AI's LFM3: a 1.7B open-weight model built for local agents, at about 1 GB in 4-bit

Liquid released six LFM3 models on October 27 (text, vision, audio, omni and two decision models), all on one 1.7B backbone. The text model has a 256K context window, runs in about 1 GB at 4-bit, and comes as GGUF, MLX and ONNX. Free commercial use stops at $10M in annual revenue.

Written by Mara at Sig · Oct 28, 2026

LFM3 is Liquid AI's new model family, out on Hugging Face as of October 27. Every model is trained from LFM3-1.7B-Base: a text model (LFM3-1.7B), a vision model that also takes video (LFM3-VL-2B), a speech model covering 18 languages (LFM3-Audio-2B), an omni model, and two "d1" models that return structured decisions from text, images or audio.

The text model is the one aimed at agents. Liquid says it was post-trained inside agent harnesses (OpenClaw, Hermes Agent and Pi), has a 256K-token context window for input plus output, and takes about 1 GB of memory at 4-bit. Checkpoints come in BF16, FP8, NVFP4, GGUF (llama.cpp, Ollama, LM Studio), MLX and ONNX. It also supports deferred tool loading, so a harness can keep most tool definitions out of the prompt until a search tool asks for them.

Liquid's own numbers: 38.42 on WebArena-Verified against 28.94 for Qwen3.5-9B, and 41.29 on the public AutomationBench against 15.10. The model card also says plainly that it isn't recommended for agentic coding or complex multi-step reasoning, and it trails Qwen3.5-9B on GPQA Diamond (63.37 vs 80.40). For hosted use, VentureBeat reports the console lists the text model at $0.08 per million input tokens and $0.20 per million output.

Read the license before you ship it. The LFM Open License v1.0 sets a commercial-use threshold at $10 million in annual revenue.

Why it matters

If your agent does one narrow job (sort tickets, pull fields out of invoices, click through one internal app), a 1 GB model you can fine-tune on a single GPU and run on a laptop changes the cost math. The catch is the license cap and Liquid's own warning that it's weak at coding and long reasoning.

Try this

Load LiquidAI/LFM3-1.7B with Transformers (trust_remote_code=True) or grab the GGUF build for llama.cpp or Ollama, use Liquid's suggested settings for tool use (temperature 0.1, top_k 50, repetition_penalty 1.1), and run 20 examples of one classification or extraction job you already have answers for.

Sources

Back to all news Get Sig Weekly every Monday