Instructions to use TirGun/Sber_Qwen3-VL-2B-Instruct-action-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use TirGun/Sber_Qwen3-VL-2B-Instruct-action-GGUF with Transformers:

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("TirGun/Sber_Qwen3-VL-2B-Instruct-action-GGUF", dtype="auto")

llama-cpp-python

How to use TirGun/Sber_Qwen3-VL-2B-Instruct-action-GGUF with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="TirGun/Sber_Qwen3-VL-2B-Instruct-action-GGUF",
	filename="Sber_Qwen3-VL-2B-Instruct-action-Q4_K_M.gguf",
)

llm.create_chat_completion(
	messages = "No input example has been defined for this model task."
)

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use TirGun/Sber_Qwen3-VL-2B-Instruct-action-GGUF with llama.cpp:

Install from brew

brew install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf TirGun/Sber_Qwen3-VL-2B-Instruct-action-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf TirGun/Sber_Qwen3-VL-2B-Instruct-action-GGUF:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf TirGun/Sber_Qwen3-VL-2B-Instruct-action-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf TirGun/Sber_Qwen3-VL-2B-Instruct-action-GGUF:Q4_K_M

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf TirGun/Sber_Qwen3-VL-2B-Instruct-action-GGUF:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf TirGun/Sber_Qwen3-VL-2B-Instruct-action-GGUF:Q4_K_M

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf TirGun/Sber_Qwen3-VL-2B-Instruct-action-GGUF:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf TirGun/Sber_Qwen3-VL-2B-Instruct-action-GGUF:Q4_K_M

Use Docker

docker model run hf.co/TirGun/Sber_Qwen3-VL-2B-Instruct-action-GGUF:Q4_K_M

LM Studio
Jan
Ollama
How to use TirGun/Sber_Qwen3-VL-2B-Instruct-action-GGUF with Ollama:
```
ollama run hf.co/TirGun/Sber_Qwen3-VL-2B-Instruct-action-GGUF:Q4_K_M
```

Unsloth Studio

How to use TirGun/Sber_Qwen3-VL-2B-Instruct-action-GGUF with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for TirGun/Sber_Qwen3-VL-2B-Instruct-action-GGUF to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for TirGun/Sber_Qwen3-VL-2B-Instruct-action-GGUF to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for TirGun/Sber_Qwen3-VL-2B-Instruct-action-GGUF to start chatting

How to use TirGun/Sber_Qwen3-VL-2B-Instruct-action-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama-server -hf TirGun/Sber_Qwen3-VL-2B-Instruct-action-GGUF:Q4_K_M

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "TirGun/Sber_Qwen3-VL-2B-Instruct-action-GGUF:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use TirGun/Sber_Qwen3-VL-2B-Instruct-action-GGUF with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama-server -hf TirGun/Sber_Qwen3-VL-2B-Instruct-action-GGUF:Q4_K_M

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default TirGun/Sber_Qwen3-VL-2B-Instruct-action-GGUF:Q4_K_M

Run Hermes

hermes

Docker Model Runner
How to use TirGun/Sber_Qwen3-VL-2B-Instruct-action-GGUF with Docker Model Runner:
```
docker model run hf.co/TirGun/Sber_Qwen3-VL-2B-Instruct-action-GGUF:Q4_K_M
```

Lemonade

How to use TirGun/Sber_Qwen3-VL-2B-Instruct-action-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull TirGun/Sber_Qwen3-VL-2B-Instruct-action-GGUF:Q4_K_M

Run and chat with the model

lemonade run user.Sber_Qwen3-VL-2B-Instruct-action-GGUF-Q4_K_M

List all available models

lemonade list

TirGun: SberRobotics Qwen3-VL-2B-Instruct-action (GGUF)

English Version | Русская версия

English Version

Lightweight and fast weights for the Qwen3-VL-2B-Instruct-action model by the SberRoboticsCenter team. This build is prepared for efficient local execution (CPU, 8GB VRAM GPU) via llama.cpp and Ollama.

Technical Information:

Conversion Tool: llama.cpp build 8373 (commit 46dba9fce).

📝 About the Model

The 2B version is a compact Vision-Language-Action (VLA) model. It is designed for tasks where low latency is critical (real-time processing) while maintaining the ability to predict coordinates for robotic systems.

This model is a specialized Vision-Language-Action (VLA) agent developed by SberRoboticsCenter. It serves as the foundational architecture and instruction-tuned base for the Green-VLA project.

The model is fine-tuned to predict precise spatial coordinates and plan robotic actions, bridging the gap between visual understanding and physical execution.

Research Paper: Green-VLA: Staged Vision-Language-Action Model for Generalist Robots

🛠 Use Cases & Practical Applications

🤖 Robotics & Control: High-precision object detection for pick-and-place tasks. The model generates coordinates [ymin, xmin, ymax, xmax] for robotic manipulators.
🖥 UI Automation & RPA: Visual navigation of software interfaces. Ideal for testing apps or automating legacy software without APIs by "seeing" buttons and fields.
🏭 Industrial Vision: Quality control on assembly lines, safety monitoring (PPE detection), and obstacle analysis for AGV/AMR robots.
📄 Advanced OCR: Extracting data from complex layouts like blueprints, receipts, and multilingual documents where context is key.

📂 Available Weights

File	Method	Size	Status
`Sber_Qwen3-VL-2B-Instruct-action-Q4_K_M.gguf`	Q4_K_M	1.03 GB	✅ Done
`Sber_Qwen3-VL-2B-Instruct-action-Q5_K_M.gguf`	Q5_K_M	1.17 GB	✅ Done
`Sber_Qwen3-VL-2B-Instruct-action-Q6_K.gguf`	Q6_K	1.32 GB	✅ Done
`Sber_Qwen3-VL-2B-Instruct-action-Q8_0.gguf`	Q8_0	1.71 GB	✅ Done
`mmproj-Sber_Qwen3-VL-2B-Instruct-action-F16.gguf`	F16	781 MB	Eyes (High)
`mmproj-Sber_Qwen3-VL-2B-Instruct-action-Q8_0.gguf`	Q8_0	424 MB	Eyes (Opt)

🚀 Usage Instructions

1. Via llama.cpp (CLI)

./llama-cli -m Sber_Qwen3-VL-2B-Instruct-action-Q4_K_M.gguf \
            --mmproj mmproj-Sber_Qwen3-VL-2B-Instruct-action-F16.gguf \
            --image "path/to/your/image.jpg" \
            -p "<|im_start|>user\n<|vision_start|><|vision_end|>Describe the objects in the photo.<|im_end|>\n<|im_start|>assistant\n" \
            -n 512 --temp 0.1

Note: In the --image parameter, specify the path to your test image (supported formats: .jpg, .png).

2. Via Ollama

Create a Modelfile:

FROM ./Sber_Qwen3-VL-2B-Instruct-action-Q4_K_M.gguf
ADAPTER ./mmproj-Sber_Qwen3-VL-2B-Instruct-action-F16.gguf
PARAMETER num_ctx 16384
PARAMETER num_gpu 99

Then: ollama create sber-action -f Modelfile

⚠️ Testing Status (WIP)

Technical correctness of the conversion is confirmed. Functional verification of Action capabilities (coordinate prediction accuracy in GGUF) is currently in progress.

⚖️ License

Distributed under Apache 2.0. Used original model by the SberRoboticsCenter team.

Developed within the TirGun project. Local AI. Privacy. Efficiency.

Русская версия

Легкие и быстрые веса модели Qwen3-VL-2B-Instruct-action от команды SberRoboticsCenter. Сборка подготовлена для эффективного запуска на локальном железе (CPU, GPU с 8GB VRAM) через llama.cpp и Ollama.

Техническая информация:

Инструмент конвертации: llama.cpp билд 8373 (коммит 46dba9fce).

📝 О модели

Версия 2B — это компактная Vision-Language-Action (VLA) модель. Она предназначена для задач, где критически важна низкая задержка (обработка в реальном времени), сохраняя при этом способность предсказывать координаты для робототехнических систем.

Данная модель является специализированным Vision-Language-Action (VLA) агентом, разработанным командой SberRoboticsCenter. Она послужила фундаментальной архитектурой и базой для проекта Green-VLA.

Модель дообучена для предсказания точных пространственных координат и планирования действий роботов, выступая связующим звеном между визуальным восприятием и физическим исполнением задач.

Научная работа: Green-VLA: Staged Vision-Language-Action Model for Generalist Robots

🛠 Области применения и практика

🤖 Робототехника: Высокоточное детектирование объектов для задач захвата (Pick-and-Place). Модель выдает нормализованные координаты [ymin, xmin, ymax, xmax] для манипуляторов.
🖥 Автоматизация UI (RPA): Визуальная навигация по интерфейсам программ. Идеально для тестирования ПО или управления старым софтом через "зрение", находя кнопки и поля ввода.
🏭 Индустриальное зрение: Контроль качества на линиях, мониторинг техники безопасности (наличие касок/жилетов) и анализ препятствий для мобильных роботов.
📄 Продвинутый OCR: Извлечение данных из чертежей, чеков и сложных мультиязычных документов, где важен контекст и структура.

📂 Состав сборки

Файл	Метод	Размер	Статус
`Sber_Qwen3-VL-2B-Instruct-action-Q4_K_M.gguf`	Q4_K_M	1.03 Гб	✅ Готов
`Sber_Qwen3-VL-2B-Instruct-action-Q5_K_M.gguf`	Q5_K_M	1.17 Гб	✅ Готов
`Sber_Qwen3-VL-2B-Instruct-action-Q6_K.gguf`	Q6_K	1.32 Гб	✅ Готов
`Sber_Qwen3-VL-2B-Instruct-action-Q8_0.gguf`	Q8_0	1.71 Гб	✅ Готов
`mmproj-Sber_Qwen3-VL-2B-Instruct-action-F16.gguf`	F16	781 Мб	Глаза (High)
`mmproj-Sber_Qwen3-VL-2B-Instruct-action-Q8_0.gguf`	Q8_0	424 Мб	Глаза (Opt)

🚀 Инструкция по запуску

1. Через llama.cpp (CLI)

./llama-cli -m Sber_Qwen3-VL-2B-Instruct-action-Q4_K_M.gguf \
            --mmproj mmproj-Sber_Qwen3-VL-2B-Instruct-action-F16.gguf \
            --image "path/to/your/image.jpg" \
            -p "<|im_start|>user\n<|vision_start|><|vision_end|>Опиши объекты на фото.<|im_end|>\n<|im_start|>assistant\n" \
            -n 512 --temp 0.1

Примечание: В параметре --image укажите путь к вашему тестовому изображению (поддерживаются форматы .jpg, .png).

2. Через Ollama

Создайте Modelfile:

FROM ./Sber_Qwen3-VL-2B-Instruct-action-Q4_K_M.gguf
ADAPTER ./mmproj-Sber_Qwen3-VL-2B-Instruct-action-F16.gguf
PARAMETER num_ctx 16384
PARAMETER num_gpu 99

Запуск: ollama create sber-action -f Modelfile

⚠️ Testing Status (WIP / В процессе)

Technical correctness of the conversion is confirmed. Functional verification of Action capabilities (coordinate prediction accuracy in GGUF) is currently in progress. Техническая корректность конвертации подтверждена. Этап верификации Action-способностей (точность предсказания координат в GGUF) находится в процессе.