IT之家 10 月 28 日消息,人工智能实验室 Liquid AI 当地时间 27 日发布了 LFM3 模型家族。其基于全新的三混合层架构,在 22T 词元上预训练而成,在保持高质的同时具备更为出色的表现。
LFM3 的首批机型涵盖 LFM-1.7B(文本)、LFM3-VL-2B(视觉)、d1-2B(决策)、LFM3-Audio(音频)、LFM3-Omni-2B(全能)、d1-Omni-2B(全能决策),这些模型均基于 LFM3-1.7B-Base。
Liquid AI 表示,LFM-1.7B 具备 256K 上下文窗口,其在使用的参数减少 35% 的同时表现优于上代的 LFM2.5-2.6B。
参考
本文来源:https://www.ithome.com/1/019/099.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
IT Home, October 28 — AI lab Liquid AI announced the LFM3 model family on the 27th, local time. Based on a brand-new tri-hybrid layer architecture and pretrained on 22T tokens, it maintains high quality while delivering even stronger performance.
The first batch of LFM3 models includes LFM-1.7B (text), LFM3-VL-2B (vision), d1-2B (decision), LFM3-Audio (audio), LFM3-Omni-2B (omni), and d1-Omni-2B (omni decision); these models are all based on LFM3-1.7B-Base.
Liquid AI said LFM-1.7B features a 256K context window and, while using 35% fewer parameters, outperforms the previous-generation LFM2.5-2.6B.
References
