LFM2.5-8B-A1B, an edge model built for fast, reliable tool calling on consumer hardware.
LFM2.5-350M & LFM2.5-8B-A1B are fast, memory-efficient hybrid MoE language models. Unsloth Dynamic (UD) quants available.
LFM2.5-8B-A1B Q8_0 with working tool calls — fixes the llama.cpp parser crash (#23838). Same weights, metadata-only fix.
Agentic coding fine-tune of LFM2.5-8B-A1B for the Klide runtime — makes correct edits, verifies, and stops instead of looping. LFM Open License v1.0.
LFM2-8B-A1B is the best on-device MoE in terms of both quality (comparable to 3-4B dense models) and speed (faster than Qwen3-1.7B). Updated LFM2.5-8B-A1B is now available!