LFM2.5-8B-A1B

LiquidAI LFM2.5-8B-A1B hybrid MoE language model for on-device text generation.

LFM2.5-8B-A1B is LiquidAI's hybrid decoder (short-conv + full-attention mixers, dense FFN on layers 0–1 and sparse MoE on 2–23). This AI Hub entry ships a refft-hexagon serving pack (LFM2.5-8B-A1B-w4.refft) with native programs for single-HTP (device_world=1) and dual-HTP (device_world=2).

Model performance

LFM2.5-8B-A1B on Hexagon HTP via refft-hexagon serving pack (1×HTP and 2×HTP as two devices).

Device

Dragonwing IQ-9075 (QCS9075) 2×HTP

SoC QCS9075 · Ubuntu 24.04.3 LTS (Noble Numbat) (aarch64) · Hexagon v73 · w4 · runtime product-local · measured 2026-08-12

Measured · QCS9075

25.3tok/s

TPS / Decode speed

557 ms – 557 ms

TTFT

23.3tok/s

Prefill speed

Install and run

Install

curl -fsSL https://raw.githubusercontent.com/refinefuture-ai/refft.cpp/main/refft-hexagon/install.sh | sh

Run

~/.local/share/refft-hexagon/bin/refft-hexagon cli --model ./LFM2.5-8B-A1B-w4.refft --backend hexagon --prompt "Who are you?" --max_new_tokens 128

Serving, environment variables and full instructions

LFM2.5-8B-A1B

Model performance

Install and run

Tags