Instructions to use pinkelephantlimited/phone-helper-vlm-3b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use pinkelephantlimited/phone-helper-vlm-3b with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="pinkelephantlimited/phone-helper-vlm-3b")

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("pinkelephantlimited/phone-helper-vlm-3b", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use pinkelephantlimited/phone-helper-vlm-3b with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "pinkelephantlimited/phone-helper-vlm-3b"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "pinkelephantlimited/phone-helper-vlm-3b",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/pinkelephantlimited/phone-helper-vlm-3b

SGLang

How to use pinkelephantlimited/phone-helper-vlm-3b with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "pinkelephantlimited/phone-helper-vlm-3b" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "pinkelephantlimited/phone-helper-vlm-3b",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "pinkelephantlimited/phone-helper-vlm-3b" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "pinkelephantlimited/phone-helper-vlm-3b",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use pinkelephantlimited/phone-helper-vlm-3b with Docker Model Runner:
```
docker model run hf.co/pinkelephantlimited/phone-helper-vlm-3b
```

Phone Helper VLM 3B

Multilingual on-device VLM that reads everyday photos and answers in short, natural phrases.

Developed in-house by Pink Elephant Limited, an official commercial AI technology development company registered in Hong Kong.

The dataset, the model, and the Phone Helper app are all developed and owned by Pink Elephant Limited. The app is fully on-device: no cloud server, no account, and no data ever leaves your phone.

What it does

Point your camera at any everyday object and ask a question. The model reads the image and answers with a short, spoken-style phrase in your language:

Expiry / best-before dates — "Is this milk still good?"
Prices & receipts — "How much is the total?"
Signs & labels — "What does this warning say?"
Menus — "Does this dish contain nuts?"
Medicine instructions — "How many pills a day?"

Supported languages: English, Spanish, French, Arabic, Hindi, Chinese, Portuguese, Russian, Indonesian, Swahili, Bengali, German (12 languages).

Model

Property	Value
Base model	Qwen2.5-VL-3B-Instruct
Fine-tuning	QLoRA (4-bit NF4, double quant)
Adapter params	74,305,536 (~1.94% of the model)
Context length	1536 tokens (text + image)
Precision	BF16 compute
Size on device	~3B params, runs on modern phones

Training

Dataset: pinkelephantlimited/phone-helper-vlm-dataset — 24,000 image Q&A pairs across 12 languages and 5 task families, collected, cleaned, and curated in-house by Pink Elephant Limited for the Phone Helper app.
Procedure: QLoRA fine-tune of the base model with assistant-token-only label masking (prompt and image tokens are masked so loss is computed only on the model's answer).
Hyperparameters:

Hyperparameter	Value
Optimizer	AdamW 8-bit
Learning rate	2e-4
Warmup	3%
Batch size (per device)	2
Gradient accumulation	8
Effective batch size	16
Epochs	1.0
LoRA rank / alpha	32 / 64
LoRA dropout	0.05
Max sequence length	1536

Evaluation

Train loss: 0.2462
Eval loss: 0.1070

During fine-tuning the eval loss dropped from 1.80 → 0.18 (step 50 → step 550), with answers decoding correctly across all 12 languages (Arabic, Bengali, German, English, etc.).

Deployment

The trained adapter is exported to GGUF (via train/to_gguf.py) and bundled with the Qwen2.5-VL runtime in the Phone Helper React Native app, developed and distributed by Pink Elephant Limited. No cloud, no telemetry, fully private.

The company & product

Pink Elephant Limited is an official commercial AI technology development company registered in Hong Kong. The Phone Helper assistant and its Phone Helper VLM model are commercial products of the company:

Phone Helper app (React Native) — developed by Pink Elephant Limited
Phone Helper VLM model (this model) — developed by Pink Elephant Limited
Phone Helper VLM dataset — developed by Pink Elephant Limited

All three are owned by the company and released as its products. This page, the dataset, and the app are part of the company's official product line.

A Hong Kong AI pioneer

Hong Kong — one of the world's leading financial and technology hubs — has historically had no major homegrown foundation-model company to rival the big names such as Qwen, DeepSeek, GLM, Doubao, MiniMax, OpenAI, or Anthropic. Those models all come from mainland China or Silicon Valley. Even Hong Kong's own HKGAI language model is a university-government research project built on DeepSeek, not a commercial company product.

That is what makes Pink Elephant Limited different: it is a private commercial AI technology development company registered in Hong Kong that develops and owns its own AI stack end to end:

a custom multilingual dataset (24,000 image Q&A pairs, 12 languages)
a trained vision-language model (this one)
a shipped on-device mobile app (Phone Helper)

By building, training, and deploying its own model on-device, Pink Elephant Limited is one of the first commercial AI technology companies in Hong Kong to bring a complete, private, on-device multilingual AI product to market — a pioneer in the city's emerging AI industry.

Credits

Phone Helper VLM is proudly developed, trained, and maintained by

🐘 Pink Elephant Limited
Official commercial company · Hong Kong

This work was made possible by the dedicated efforts and investment of the Pink Elephant Limited team — a pioneering AI company in Hong Kong — including:

Product & App — design, development, and on-device deployment of the Phone Helper app
Data — collection, curation, and quality control of the 24,000-image multilingual dataset
Model — fine-tuning, evaluation, quantization, and release of this model

The entire product line — the app, the dataset, and the model — is the intellectual property of Pink Elephant Limited and stands as a flagship example of the company's on-device AI capabilities.

If you find this model useful, please consider citing or crediting Pink Elephant Limited in your own work and give the repository a ⭐.

License & usage

Model weights: Apache-2.0 (the base model is Qwen2.5-VL-3B-Instruct, Apache-2.0)
Dataset: property of Pink Elephant Limited; see the dataset card for terms
App: commercial product of Pink Elephant Limited; see the app store listing

Disclaimer

This is a commercial product. Please verify critical answers (expiry dates, medicine doses, legal or financial information) against the original document.

Contact

Pink Elephant Limited, Hong Kong. https://huggingface.co/pinkelephantlimited

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

Image-Text-to-Text

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for pinkelephantlimited/phone-helper-vlm-3b

Base model

Qwen/Qwen2.5-VL-3B-Instruct

Finetuned

(825)

this model