Liquid AI's LFM3-1.7B base model was pretrained on 22 trillion tokens, with the final 3.5 trillion tokens dedicated to agentic mid-training to extend its context window to 256,000 tokens