TL;DR: Liquid AI's largest LFM2.5 model packs 24B total / 2B active parameters with 131K context, targeting on-device agentic work in laptops and PCs.
Summary: Liquid AI released LFM2.5-24B-A2B, the largest member of its LFM2.5 family, a sparse mixture-of-experts reasoning model with 24B total parameters and 2B active per token. It ships with a 131K-token context window, a 128K-token vocabulary, and was pretrained on 33T tokens, with open base and post-trained weights published on Hugging Face. Liquid claims it performs competitively with Qwen3.5-35B-A3B — a model with roughly 50% more total and active parameters — on agentic and tool-use benchmarks, and positions it as easy to fine-tune for local agent harnesses.
Why it matters: A 2B-active model that holds up on tool-use tasks makes always-on local agents viable without cloud inference costs. Worth benchmarking against Qwen3.5-35B-A3B in your own harness and testing LoRA fine-tunes on the open weights.
Source: x_com