Liquid: LFM 7B

LFM-7B, released on January 25, 2025, is a 7B parameter language model from Liquid, featuring a substantial 32768 context length. The model consistently ranks among the fastest and most competitively priced models across various benchmarks. In terms of performance, LFM-7B demonstrates a notable strength in hallucination mitigation, achieving 86.0% accuracy and a best accuracy-to-cost ratio, indicating its ability to appropriately acknowledge uncertainty. It also performs well in Email Classification with 95.0% accuracy, standing out as the most accurate model at its price point. The model shows promising results in Reasoning, with a 36.0% accuracy and the best accuracy-to-cost ratio in this category. However, LFM-7B exhibits significant weaknesses in General Knowledge, Ethics, and Coding, where its accuracy scores are very low or 0.0%. Instruction Following also presents a challenge with 29.0% accuracy. Despite these areas for improvement, its low memory footprint and fast inference speed, coupled with its competitive pricing, position it as a cost-effective option for chat capabilities, particularly in English, Arabic, and Japanese.