LFM2.5-VL-3B - Liquid AI Docs

← Back to Vision Models LFM2.5-VL-3B is Liquid AI’s most capable vision-language model, with grounding, screen and document understanding, and function calling. It builds on LFM2-VL-3B with further mid- and post-training, pairs the LFM2.5-2.6B language backbone with a SigLIP2 NaFlex shape-optimized 400M vision encoder, and answers directly without a reasoning trace for low-latency inference on-device and in the cloud.

HF GGUF MLX ONNX DSpark

Specifications

Property	Value
Parameters	3B
Context Length	32K tokens
Architecture	LFM2.5-VL (Dense)
Languages	English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Spanish, Vietnamese, Thai, Indonesian, Hindi, Russian, Polish
License	LFM Open License v1.0

Grounding & Detection

Object detection and localization from natural-language queries

Screen & Document Understanding

Digital screens, full-page OCR, and layout-aware parsing

Function Calling

Tool use from text-only and vision-text inputs

Quick Start

Transformers
vLLM
SGLang
llama.cpp

Speculative Decoding

SGLang
llama.cpp
MLX-VLM

Install:

pip install "mlx-vlm>=0.7.2"

Run with the drafter:

python -m mlx_vlm.generate \
    --model LiquidAI/LFM2.5-VL-3B \
    --draft-model LiquidAI/LFM2.5-VL-3B-DSpark \
    --draft-block-size 8 \
    --image /path/to/image.jpg \
    --prompt "Describe this image." \
    --max-tokens 256 \
    --temperature 0

Flags:

--draft-model: the drafter repo on Hugging Face. MLX-VLM detects that it is a DSpark drafter.
--draft-block-size: how many tokens the drafter proposes per step. Use 8 on Apple silicon.
--temperature 0: DSpark in MLX-VLM currently supports greedy sampling only.

​Specifications