LFM2.5-VL-3B

LiquidAI LFM2.5-VL-3B hybrid decoder + SigLIP2 (text-first Hexagon pack).

LFM2.5-VL-3B is LiquidAI's 3B vision-language model. The language backbone is the same 30-layer hybrid decoder as LFM2.5-2.6B (22 double-gated short-conv mixers + 8 GQA layers, dense SwiGLU), with rope_theta 1e6 and HF keys under model.language_model.*. Vision is SigLIP2-so400m + gelu projector (not in the first text pack). The hexagon package is a host VISA client (lfm2_5_vl_3b) talking to refft-hexagon.

Model performance

LFM2.5-VL-3B text-first on Hexagon HTP via refft-hexagon cli --model.

Snapdragon 8 Elite Gen 5 (SM8850) 1×HTP

SoC SM8850 · Android (aarch64) · Hexagon v81 · w4 · runtime 0.6.5.dev691+gdf361427b · measured 2026-08-13

Measured · SM8850

33.4tok/s

TPS / Decode speed

117 ms

TTFT

159tok/s

Prefill speed

Install and run

Install

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

Run

adb shell /data/local/tmp/refft-hexagon/bin/refft-hexagon cli --model /data/local/tmp/LFM2.5-VL-3B-w4.refft --backend hexagon --prompt "Who are you?" --max_new_tokens 128

Serving, environment variables and full instructions

LFM2.5-VL-3B

Model performance

Install and run

Tags