---

base_model: Qwen/Qwen2-VL-2B-Instruct

license: apache-2.0

library_name: gguf

pipeline_tag: image-text-to-text

tags:

- gguf

- qwen2-vl

- vision

- multilingual

- on-device

- llama.cpp

language: [en, es, fr, ar, hi, zh, pt, ru, id, sw, bn, de]

---

Pink Elephant Talk — VLM (v3.0, Qwen2-VL-2B, GGUF)

GGUF quantized release used by Pink Elephant Talk,

the on-device vision-language personal assistant from

Pink Elephant Limited (Hong Kong).

v3.0 ships the smaller, faster Qwen2-VL-2B-Instruct base for quicker on-device

inference on mobile hardware. These files power offline inference with

llama.cpp / llama.rn.

Important licensing note

• Base model: This is a GGUF conversion of Qwen2-VL-2B-Instruct,

© Alibaba Group, released under the Apache License, Version 2.0.

Apache 2.0 requires retaining the license and NOTICE when redistributing. Full text:

<https://www.apache.org/licenses/LICENSE-2.0>

• Conversion: Quantized GGUF + vision projector (mmproj) derived from the

community conversion by bartowski (https://huggingface.co/bartowski/Qwen2-VL-2B-Instruct-GGUF),

used here under the same Apache 2.0 terms.

• Pink Elephant's portion: The application and the Pink Elephant datasets are

owned by Pink Elephant Limited. Pink Elephant does not claim the Qwen base model as

its own work; it redistributes it under Apache 2.0. "Qwen" is a trademark of Alibaba.

Files

| File | Size | Purpose |

| --- | --- | --- |

| phone-helper-2b-q4_0.gguf | ~0.9 GB | Qwen2-VL-2B text+vision model (Q4_0) — recommended, GPU (Adreno/OpenCL) compatible |

| phone-helper-2b.f16-mmproj.gguf | ~1.3 GB | Vision projector (mmproj), f16 — pairs with the Q4_0 model |

Legacy (v2.0, Qwen2.5-VL-3B) files remain in this repo for backward compatibility

with the v2.0 build: phone-helper-3b-q4_0.gguf, phone-helper-3b-q4_0.f16-mmproj.gguf,

phone-helper-3b-q4_k_m.gguf, phone-helper-3b-q4_k_m.f16-mmproj.gguf.

GPU note: llama.rn / llama.cpp GPU acceleration on Android (OpenCL / Adreno) currently

supports Q4_0 and Q6_K quant formats only — use the Q4_0 pair for on-device GPU

inference and load both the .gguf model and its paired .mmproj file together.

Usage (llama.cpp)

llama-cli -m phone-helper-2b-q4_0.gguf \
          -mm phone-helper-2b.f16-mmproj.gguf \
          -p "What is the expiry date on this?" \
          --image photo.jpg

Python (llama-cpp-python)

from llama_cpp import Llama
llm = Llama(model_path="phone-helper-2b-q4_0.gguf",
            mmproj="phone-helper-2b.f16-mmproj.gguf")

On-device app (Pink Elephant Talk)

This model is bundled in the Pink Elephant Talk app by Pink Elephant Limited

(Android + iOS). The vision+language model runs entirely on the device — your

prompts and photos are not sent to any cloud LLM API (no hosted inference). The app

downloads the model once on first launch; it may optionally fetch current web data

(news / facts) to answer live questions, and that optional fetch goes directly to public

sources such as Google News RSS and Wikipedia — never to an LLM cloud.

Recommended device: Android with 8 GB RAM or more (~3 GB free for the model).

About the company

Pink Elephant Limited is a commercial

AI technology development company registered in Hong Kong. It develops the datasets and

the app in-house. Base model attribution to Alibaba Group / Qwen and to the GGUF

conversion author is preserved per the Apache 2.0 license; Pink Elephant claims its own

application and datasets only.