LFM2.5-24B-A2B

LiquidAI released LFM2.5-24B-A2B, a 23.97B-total, 2B-active mixture-of-experts model tuned for desktop-agent workloads, with weights posted October 9 and refreshed October 12 with an FP8 build. The model ships with a 131,072-token context, official GGUF, FP8, Base and ONNX builds, and a bespoke Liquid AI license; the GGUF drew 17,807 downloads in days while the model page logged 4,067 downloads and 234 likes. Its 13.7GB of q4 weights and 40-layer cache at 80MB per 1,000 tokens let a 24GB card run it with a real context window and a 32GB box run it with headroom for the tool loop.

The always-on-agent MoE. LiquidAI’s LFM2.5-24B-A2B weights October 9, refreshed October 12 with the FP8 build is a 23.97B-total, 2B-active mixture of experts tuned for the desktop-agent workloads the d1 edge line teases: 131,072-token context, official GGUF out the same day 17,807 downloads in days , plus FP8, Base and ONNX builds. Fit math. 13.7GB of q4 weights and a 40-layer cache at 80MB per 1,000 tokens 8 KV heads x 64 head-dim : a 24GB card runs it with a real context window, a 32GB class box runs it with headroom for the tool loop. The 2B active slice means per-token compute tracks a small model while 24B of knowledge stays resident, which is exactly the trade an always-on harness wants. The license, again bespoke. Same Liquid AI license terms as the d1 line: downloadable, deployable, with fleet-scale terms to read before embedding. 4,067 downloads and 234 likes in the first days; the GGUF carries 17,807 more.