LFM2.5-24B-A2B (2.68bit) matches Qwen3.8-27B on monocoder test

The test is part of a running series that measures locally runnable models under a fixed set of prompts, with the same Mac mini, the same llama-server engine, and the same 32768-token context. LFM2.5-24B-A2B is a mixture-of-experts model from Liquid AI with 24B total parameters but only about 2.45B active per token, per the official model card, and the author chose a roughly 2.68-bit quantization to fit it onto 16GB of memory.

A key complication is that LFM2.5 always thinks before answering and cannot be told to stop; a client request to disable thinking had no effect. That made the first monocoder run produce no usable answer within the response cap, which the author treated as a harness problem rather than a model failure, adding a guard to exclude empty capped responses and rerunning LFM2.5 alone with a 16384-token cap. The author also notes that quantization effects could not be separated, since no higher-bit version of the same model was tested as a control, and that official speed or performance figures were not checked in this environment.