LFM2-24B-A2B: Scaling Up the LFM2 Architecture
Summary
Liquid AI released LFM2-24B-A2B, a sparse Mixture of Experts model with 24B total parameters and 2B active per forward pass, designed to run on 32GB RAM and enable edge deployments. The post details the architecture, throughput benchmarks, and open-weight availability, highlighting edge-friendly inference via llama.cpp, vLLM, and SGLang, plus future on-device NPU support. It also emphasizes accessible access through Hugging Face and Playground for developers.