• Sources: primary, discussion
  • Summary: LFM2.5-24B-A2B keeps the LFM2 hybrid architecture of gated short convolutions and GQA with 64 experts and top-4 routing. It raises context from 32K to 131K tokens and was pre-trained on about 33T tokens. Liquid AI reports 184 tokens per second on an Apple M5 Max and 51.0 on SWE-Bench Verified, below Qwen3.6-35B-A3B at 66.2. Base and instruct weights are on Hugging Face.
  • Why it matters: It is an open-weight option for local tool-calling agents on laptop-class hardware, but its coding scores trail larger MoEs.