Liquid AI releases LFM3: 1.7B small model outperforms 9B in running web agents.

Beating AI News Flash: Liquid AI has released its third-generation foundation model LFM3, unveiling six small models with approximately 2 billion parameters at once, covering text, vision, speech, and decision-making tasks. Among them, the 1.7B text model supports a 256K context, and its 4-bit quantized weights are only about 1GB. All six models support fine-tuning on a single GPU, allowing enterprises to train them specifically with their own data before deploying them to phones, computers, or private servers.

Liquid AI redesigned the model architecture, combining short convolutions, local attention, and global attention, and conducted specialized training for Agent tasks such as tool calling and web operations. Compared with the previous generation LFM2.5-2.6B, the new model reduces parameters by about 35%, while achieving substantial improvements on some tasks.

In official tests, LFM3-1.7B scored 38.42 on the WebArena-Verified browser operation task, compared with only 3.02 for the previous generation, and it also surpassed Qwen3.5-9B (28.94), which has about 5 times its parameter count. In the AutomationBench enterprise workflow test, its task success rate reached 41.29%, while Qwen3.5-9B scored 15.10%.

But when switching to complex programming and long-text tasks, the situation reversed. In the SWE-bench Verified programming test, LFM3-1.7B scored 43.4, lower than Qwen3.5-9B's 51.2. In the 128K long-text test, the two scored 65.6 and 87.1, respectively. Liquid AI also explicitly stated in the model card that this small model is not recommended for complex Agent programming tasks.