Skip to content
Computer vision

Pony-AndroidGrounding

2B-модель для поиска элементов на скриншотах Android

2 min read

Pony-AndroidGrounding — vision-language модель от JivinDotL, дообученная для visual grounding по инструкции на скриншотах Android. База — Qwen3-VL-2B-Instruct (2,13 млрд параметров): LoRA r=16 на q/k/v/o_proj, затем merge_and_unload в полные bf16-веса. В обучение также входит кастомный корпус Identity + Behavior SFT + DPO.

Ключевые особенности

  • На инструкцию возвращает ровно один bbox в нормализованных координатах 0–1000; если цели нет — 未找到.
  • Tool-call протокол (<tool_call> JSON; locate_ui_element / tap) — 6/6 под строгим парсером; модель только предлагает координаты и не выполняет действия.
  • На тесте (115): IoU50 80.4, IoU75 72.5, click 87.5; корректный отказ и валидный JSON — по 100.
  • Dense enumeration holdout (52): IoU50 100.0, IoU75 98.1.
  • ~0.5 с на запрос в bf16 на одной датацентровой GPU; веса Apache-2.0.

Пример

from transformers import AutoModelForImageTextToText, AutoProcessor
model = AutoModelForImageTextToText.from_pretrained("JivinDotL/pony-android-grounding", dtype="bfloat16")
proc = AutoProcessor.from_pretrained("JivinDotL/pony-android-grounding")
# instruction + portrait screenshot -> <bbox>x1,y1,x2,y2</bbox>
# если цели нет — 未找到
ПодробнееТест (115, отложенный домен приложений): IoU50 80.4, IoU75 72.5, click 87.5, корректный отказ 100, валидный JSON 100.
Val (225): IoU50 87.3, IoU75 85.0, click 93.5.
Латентность ~0.5 с (bf16, одна датацентровая GPU); обучение — портретные 720×1600, landscape/web вне OOD.
Слабые места: иконки <40 px и близкие группы иконок; Q6-проба про «Pony.ai» склонна к конфабуляции.
Веса Apache-2.0; скриншоты обучения не распространяются.
Discuss Telegram 1 5

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 4.0 GB · ready files4.9 GB
  • FP8 / INT8recommendedweights 2.0 GB2.7 GB
  • 4-bitweights 1.1 GB1.7 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · BF16 — from RTX 2060

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # JivinDotL/pony-android-grounding
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download JivinDotL/pony-android-grounding --local-dir ~/models/pony-android-grounding
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to pony-android-grounding
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of Qwen3-VL-2B-Instruct (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 2All on Hugging Face
Merges 1All on Hugging Face

About the model

JivinDotL/pony-android-grounding
Author
JivinDotL
Category
Computer vision
Released
Posted

On NeyroLearn

Hugging Face

51

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments