Liquid AI
United StatesModel developer
MIT spinoff building non-Transformer Liquid Foundation Models.
What they do
Liquid AI's LFM series (LFM2, LFM2.5, including tiny 1.2B 'Thinking' and vision-language variants) uses architectures inspired by liquid neural networks rather than pure Transformers, targeting on-device efficiency. Backed heavily by AMD.
Models they develop
9 in this directory.
- LFM 40BLiquid AIOpen40B parametersResearch
- LFM2 1.2BLiquid AIOpen1.2B parametersResearch
- LFM2 2.6BLiquid AIOpen2.6B parametersResearch
- LFM2 24B A2BLiquid AIOpen24B total / 2B active (MoE)Research
- LFM2 8B A1BLiquid AIOpen8B total / 1B active (MoE)Research
- LFM2.5-1.2B-InstructLiquid AIOpen1.2B parametersResearch
- LFM2.5-1.2B-ThinkingLiquid AIOpen1.2B parametersResearch
- LFM2.5-8B-A1BLiquid AIOpen8B total / 1B active (MoE)Research
- LFM2.5-VL-1.6BLiquid AIOpen1.6B parametersResearch