Article
Liquid AI Release LFM3，1.7B Small models run webpages Agent outperform 9B Dongcha Beating AI breaking news, Liquid AI releases its third-generation foundation models LFM3, launching six small models of approximately 2000000000 parameters at once, covering text, vision, speech, and decision-making tasks. Among them, 1.7B the text model supports 256K context, 4-bit and its quantized weights are only approximately 1GB. All six models support single-GPU fine-tuning, allowing companies to use their own data for specialized training and then deploy the models on phones, computers, or private servers.
Liquid AI redesigned the model architecture by combining short convolution, local attention, and global attention, and specifically trained it for Agent tasks such as tool calling and webpage operation. Compared with the previous-generation LFM2.5-2.6B, the new models have approximately 35%fewer parameters, yet their performance improved substantially on some tasks.
In official tests,LFM3-1.7B scored 38.42 on the WebArena-Verified browser operation task, while the previous generation scored only 3.02, and it also surpassed 5), whose parameter count is approximately Qwen3.5-9B（28.94 times larger. In the AutomationBench enterprise workflow test, its task success rate reached 41.29%，Qwen3.5-9B was 15.10%.
However, the results reversed on complex programming and long-text tasks. In the SWE-bench Verified programming test,LFM3-1.7B scored 43.4, lower than Qwen3.5-9B's 51.2。128K long-text test, the two scores were 65.6 and 87.1。Liquid AI also explicitly states in its model card that it does not recommend using this small model for complex Agent programming tasks.
Original link https://m.theblockbeats.info/flash/374160