According to CoinMeta, Liquid AI has released its third-generation foundation model, LFM3, which includes six smaller models with approximately 2 billion parameters each, covering text, vision, speech, and decision-making tasks. Among them, the 1.7B text model supports a context of 256K, and the weights after 4-bit quantization are only about 1GB. All six models support fine-tuning on a single GPU, allowing for specialized training using a company's own data before deployment on mobile phones, computers, or private servers. Compared to the previous generation LFM2.5-2.6B, the new models have about 35% fewer parameters, yet their performance in some tasks has significantly improved. In official tests, LFM3-1.7B scored 38.42 on WebArena-verified browser operation tasks, while the previous generation scored only 3.02; it also surpassed Qwen3.5-9B (with about 5 times the number of parameters) with a score of 28.94. In AutomationBench enterprise workflow tests, its task success rate reached 41.29%, compared to 15.10% for the previous generation. However, in complex programming and long-text tasks, the performance of LFM3-1.7B has declined.