The short version
A 2.6B-parameter local model aimed at lightweight agentic and extraction workloads.
From the HackerLinks archive
LFM2.5 2.6B
A 2.6B-parameter local model aimed at lightweight agentic and extraction workloads.
At a glance:
First seen:2026-08-12
Last seen:2026-08-12
Times seen:1
Website:huggingface.co
Why it caught our attention
It has concrete evidence of useful local speed on unusually old hardware, alongside a case for intentionally small-model design.
Where it surfaced on Hacker News
2026-08-12
“the official 6-bit quant of 2.6B runs at 25-30 tok/s under llama.cpp on one of the D500 GPUs.”
h14h · recommendation · first hand use
Direct HN comment
“They target reliable operation of tiny models in ways other model families don't”
0xbadcafebee · comparison
Direct HN comment
LFM2.5 2.6B model competitive with 4x larger models