LFM2.5-24B-A2B
Liquid AI · 2026-10-13 · 24B parameters
LFM2.5-24B-A2B is Liquid AI's post-trained sparse MoE reasoning model, released with open weights on October 13, 2026 as the largest and final LFM2.5 model and distinct from LFM2-24B-A2B. It keeps the 40-layer hybrid of 30 double-gated short-convolution blocks and 10 GQA blocks with 64 experts and top-4 routing, extends context from 32,768 to 131,072 tokens, and expands the tokenizer; Liquid reports 2.45B active parameters. Unlike its predecessor it always thinks before answering, and it was trained with agentic reinforcement learning inside OpenClaw, Hermes Agent, Pi and OpenCode. The BF16 checkpoint contains 23,971,587,712 parameters; the tokenizer assigns 125,017 distinct IDs, while the configuration and card advertise 128,000 embedding rows.
Benchmark scores
| Benchmark | Score |
|---|
No benchmark scores recorded.