LFM2-24B-A2B
Liquid AI · 2026-02-24 · 23.8B parameters
LFM2 24B A2B is Liquid AI's larger sparse MoE model with 24 billion total parameters and 2.3 billion active per token, scaling up the LFM2 hybrid architecture for stronger capability while remaining deployable on consumer hardware with 32 GB RAM. It extends the LFM2 design—pairing gated short-convolution blocks with GQA blocks—going to 40 layers and 64 experts per MoE block with top-4 routing, up from the 24-layer, 32-expert LFM2-8B-A1B. Released as open weights in February 2026, it supports llama.cpp, vLLM, and SGLang inference with multiple quantization options.