Model Overview
Model Features
Model Capabilities
Use Cases
๐ Qwen2.5-VL-3B-Instruct GGUF Models
These files are built using an imatrix file and the latest llama.cpp build. To use the vision feature with the model, you must use a fork of llama.cpp.
๐ Quick Start
โจ Features
- These GGUF models of Qwen2.5-VL-3B-Instruct are built with specific techniques, enabling vision functionality with a llama.cpp fork.
- The latest quantization method introduces precision - adaptive quantization, proven effective on Llama - 3 - 8B, with layer - specific strategies for accuracy and memory efficiency.
- Multiple model formats are available to suit different hardware capabilities and memory constraints.
๐ฆ Installation
How to Use Qwen 2.5 VL Instruct with llama.cpp
To use the experimental support for Qwen 2.5 VL in llama.cpp
, follow these steps. Note that this uses a fork of llama.cpp, and currently, the main branch does not support vision for this model.
- Clone the latest llama.cpp Fork:
git clone https://github.com/HimariO/llama.cpp.qwen2vl.git cd llama.cpp.qwen2vl git checkout qwen25 - vl - 20250404
- Build the Llama.cpp:
Build llama.cpp as usual: https://github.com/ggml-org/llama.cpp#building-the-project.
Once llama.cpp is built, copy the
./llama.cpp.qwen2vl/build/bin/llama-qwen2-vl-cli
to a chosen folder. - Download the Qwen 2.5 VL gguf file:
https://huggingface.co/Mungert/Qwen2.5-VL-3B-Instruct-GGUF/tree/main
Choose a gguf file without the
mmproj
in the name. Example gguf file: https://huggingface.co/Mungert/Mungert/Qwen2.5-VL-3B-Instruct-GGUF/resolve/main/Qwen2.5-VL-3B-Instruct-q8_0.gguf Copy this file to your chosen folder. - Download the Qwen 2.5 VL mmproj file
https://huggingface.co/Mungert/Qwen2.5-VL-3B-Instruct-GGUF/tree/main
Choose a file with
mmproj
in the name. Example mmproj file: https://huggingface.co/Mungert/Qwen2.5-VL-3B-Instruct-GGUF/resolve/main/Qwen2.5-VL-3B-Instruct-mmproj-f16.gguf Copy this file to your chosen folder. - Copy images to the same folder as the gguf files or alter paths appropriately.
In the example below, the gguf files, images, and
llama-qwen2vl-cli
are in the same folder. Example image: image Copy this file to your chosen folder.
๐ป Usage Examples
Basic Usage
llama-qwen2vl-cli -m Qwen2.5-VL-3B-Instruct-q8_0.gguf --mmproj Qwen2.5-VL-3B-Instruct-mmproj-f16.gguf -p "Describe this image." --image ./car-1.jpg
Advanced Usage
Quick test with the highest of the DynamicGate quants IQ2_M
on Qwen 2.5 VL 3B:
llama.cpp.qwen2vl/build/bin/llama-qwen2vl-cli -m Qwen2.5-VL-3B-Instruct-iq2_m.gguf --mmproj qwen.qwen2.5-vl-3b-instruct-vision.f16.gguf -p "Describe this image in a lot of detail." --image ./car-1.jpg
๐ง Technical Details
Ultra - Low - Bit Quantization with IQ - DynamicGate (1 - 2 bit)
Our latest quantization method introduces precision - adaptive quantization for ultra - low - bit models (1 - 2 bit), with benchmark - proven improvements on Llama - 3 - 8B. This approach uses layer - specific strategies to preserve accuracy while maintaining extreme memory efficiency.
Benchmark Context
All tests were conducted on Llama - 3 - 8B - Instruct using:
- Standard perplexity evaluation pipeline
- 2048 - token context window
- The same prompt set across all quantizations
Method
- Dynamic Precision Allocation:
- First/Last 25% of layers โ IQ4_XS (selected layers)
- Middle 50% โ IQ2_XXS/IQ3_S (increase efficiency)
- Critical Component Protection:
- Embeddings/output layers use Q5_K
- Reduces error propagation by 38% vs standard 1 - 2bit
Quantization Performance Comparison (Llama - 3 - 8B)
Quantization | Standard PPL | DynamicGate PPL | ฮ PPL | Std Size | DG Size | ฮ Size | Std Speed | DG Speed |
---|---|---|---|---|---|---|---|---|
IQ2_XXS | 11.30 | 9.84 | -12.9% | 2.5G | 2.6G | +0.1G | 234s | 246s |
IQ2_XS | 11.72 | 11.63 | -0.8% | 2.7G | 2.8G | +0.1G | 242s | 246s |
IQ2_S | 14.31 | 9.02 | -36.9% | 2.7G | 2.9G | +0.2G | 238s | 244s |
IQ1_M | 27.46 | 15.41 | -43.9% | 2.2G | 2.5G | +0.3G | 206s | 212s |
IQ1_S | 53.07 | 32.00 | -39.7% | 2.1G | 2.4G | +0.3G | 184s | 209s |
Key:
- PPL = Perplexity (lower is better)
- ฮ PPL = Percentage change from standard to DynamicGate
- Speed = Inference time (CPU avx2, 2048 token context)
- Size differences reflect mixed quantization overhead
Key Improvements:
- ๐ฅ IQ1_M shows a massive 43.9% perplexity reduction (27.46 โ 15.41)
- ๐ IQ2_S cuts perplexity by 36.9% while adding only 0.2GB
- โก IQ1_S maintains 39.7% better accuracy despite 1 - bit quantization
Tradeoffs:
- All variants have modest size increases (0.1 - 0.3GB)
- Inference speeds remain comparable (<5% difference)
When to Use These Models
๐ Fitting models into GPU VRAM โ Memory - constrained deployments โ Cpu and Edge Devices where 1 - 2bit errors can be tolerated โ Research into ultra - low - bit quantization
๐ Documentation
Choosing the Right Model Format
Selecting the correct model format depends on your hardware capabilities and memory constraints.
Property | Details |
---|---|
BF16 (Brain Float 16) โ Use if BF16 acceleration is available | A 16 - bit floating - point format designed for faster computation while retaining good precision. Provides similar dynamic range as FP32 but with lower memory usage. Recommended if your hardware supports BF16 acceleration (check your deviceโs specs). Ideal for high - performance inference with reduced memory footprint compared to FP32. |
F16 (Float 16) โ More widely supported than BF16 | A 16 - bit floating - point format with high precision but a smaller range of values than BF16. Works on most devices with FP16 acceleration support (including many GPUs and some CPUs). Slightly lower numerical precision than BF16 but generally sufficient for inference. |
Quantized Models (Q4_K, Q6_K, Q8, etc.) โ For CPU & Low - VRAM Inference | Quantization reduces model size and memory usage while maintaining as much accuracy as possible. Lower - bit models (Q4_K) are best for minimal memory usage but may have lower precision. Higher - bit models (Q6_K, Q8_0) offer better accuracy but require more memory. |
Very Low - Bit Quantization (IQ3_XS, IQ3_S, IQ3_M, Q4_K, Q4_0) | These models are optimized for extreme memory efficiency, making them ideal for low - power devices or large - scale deployments where memory is a critical constraint. |
Included Files & Details
File Name | Details |
---|---|
Qwen2.5-VL-3B-Instruct-bf16.gguf |
Model weights preserved in BF16. Use this if you want to requantize the model into a different format. Best if your device supports BF16 acceleration. |
Qwen2.5-VL-3B-Instruct-f16.gguf |
Model weights stored in F16. Use if your device supports FP16, especially if BF16 is not available. |
Qwen2.5-VL-3B-Instruct-bf16-q8_0.gguf |
Output & embeddings remain in BF16. All other layers quantized to Q8_0. Use if your device supports BF16 and you want a quantized version. |
Qwen2.5-VL-3B-Instruct-f16-q8_0.gguf |
Output & embeddings remain in F16. All other layers quantized to Q8_0. |
Qwen2.5-VL-3B-Instruct-q4_k.gguf |
Output & embeddings quantized to Q8_0. All other layers quantized to Q4_K. Good for CPU inference with limited memory. |
Qwen2.5-VL-3B-Instruct-q4_k_s.gguf |
Smallest Q4_K variant, using less memory at the cost of accuracy. Best for very low - memory setups. |
Qwen2.5-VL-3B-Instruct-q6_k.gguf |
Output & embeddings quantized to Q8_0. All other layers quantized to Q6_K. |
Qwen2.5-VL-3B-Instruct-q8_0.gguf |
Fully Q8 quantized model for better accuracy. Requires more memory but offers higher precision. |
Qwen2.5-VL-3B-Instruct-iq3_xs.gguf |
IQ3_XS quantization, optimized for extreme memory efficiency. Best for ultra - low - memory devices. |
Qwen2.5-VL-3B-Instruct-iq3_m.gguf |
IQ3_M quantization, offering a medium block size for better accuracy. Suitable for low - memory devices. |
Qwen2.5-VL-3B-Instruct-q4_0.gguf |
Pure Q4_0 quantization, optimized for ARM devices. Best for low - memory environments. Prefer IQ4_NL for better accuracy. |
๐ License
The license for this project is qwen - research.
๐ If you find these models useful
Please click like โค. Also, Iโd really appreciate it if you could test my Network Monitor Assistant at ๐ Network Monitor Assitant. ๐ฌ Click the chat icon (bottom right of the main and dashboard pages). Choose an LLM; toggle between the LLM Types TurboLLM -> FreeLLM -> TestLLM.
What I'm Testing
I'm experimenting with function calling against my network monitoring service using small open - source models. I'm exploring the question "How small can it go and still function". ๐ก TestLLM โ Runs the current testing model using llama.cpp on 6 threads of a Cpu VM (Should take about 15s to load. Inference speed is quite slow and it only processes one user prompt at a timeโstill working on scaling!). If you're curious, I'd be happy to share how it works!
The other Available AI Assistants
๐ข TurboLLM โ Uses gpt - 4o - mini. Fast! Note: tokens are limited since OpenAI models are pricey, but you can Login or Download the Free Network Monitor agent to get more tokens. Alternatively, use the TestLLM. ๐ต HugLLM โ Runs open - source Hugging Face models. Fast, runs small models (โ8B) hence lower quality. Get 2x more tokens (subject to Hugging Face API availability)
Introduction
In the past five months since Qwen2 - VLโs release, numerous developers have built new models on the Qwen2 - VL vision - language models, providing us with valuable feedback. During this period, we focused on building more useful vision - language models. Today, we are excited to introduce the latest addition to the Qwen family: Qwen2.5 - VL.
Key Enhancements
- Understand things visually: Qwen2.5 - VL is not only proficient in recognizing common objects such as flowers, birds, fish, and insects but is highly capable of analyzing texts, charts, icons, graphics, and layouts within images.
- Being agentic: Qwen2.5 - VL directly plays as a visual agent that can reason and dynamically direct tools, which is capable of computer use and phone use.
- Understanding long videos and capturing events: Qwen2.5 - VL can comprehend videos of over 1 hour, and this time it has a new ability of capturing events by pinpointing the relevant video segments.
- Capable of visual localization in different formats: Qwen2.5 - VL can accurately localize objects in an image by generating bounding boxes or points and can provide stable JSON outputs for coordinates and attributes.
- Generating structured outputs: for data like scans of invoices, forms, tables, etc. Qwen2.5 - VL supports structured outputs of their contents, benefiting usages in finance, commerce, etc.
Model Architecture Updates
- Dynamic Resolution and Frame Rate Training for Video Understanding: We extend dynamic resolution to the temporal dimension by adopting dynamic FPS sampling, enabling the model to comprehend videos at various sampling rates. Accordingly, we update mRoPE in the time dimension with IDs and absolute time alignment, enabling the model to learn temporal sequence and speed and ultimately acquire the ability to pinpoint specific moments.
- Streamlined and Efficient Vision Encoder We enhance both training and inference speeds by strategically implementing window attention into the ViT. The ViT architecture is further optimized with SwiGLU and RMSNorm, aligning it with the structure of the Qwen2.5 LLM.
We have three models with 3, 7, and 72 billion parameters. This repo contains the instruction - tuned 3B Qwen2.5 - VL model. For more information, visit our [Blog](https://qwenlm.github.io/blog/qwen2.5 - vl/) and [GitHub](https://github.com/QwenLM/Qwen2.5 - VL).
Evaluation
Image benchmark
Benchmark | InternVL2.5 - 4B | Qwen2 - VL - 7B | Qwen2.5 - VL - 3B |
---|---|---|---|
MMMUval | 52.3 | 54.1 | 53.1 |
MMMU - Proval | 32.7 | 30.5 | 31.6 |
AI2Dtest | 81.4 | 83.0 | 81.5 |
DocVQAtest | 91.6 | 94.5 | 93.9 |
InfoVQAtest | 72.1 | 76.5 | 77.1 |
TextVQAval | 76.8 | 84.3 | 79.3 |
MMBench - V1.1test | 79.3 | 80.7 | 77.6 |
MMStar | 58.3 | 60.7 | 55.9 |
MathVistatestmini | 60.5 | 58.2 | 62.3 |
MathVisionfull | 20.9 | 16.3 | 21.2 |
Video benchmark
Benchmark | InternVL2.5 - 4B | Qwen2 - VL - 7B | Qwen2.5 - VL - 3B |
---|---|---|---|
MVBench | 71.6 | 67.0 | 67.0 |
VideoMME | 63.6/62.3 | 69.0/63.3 | 67.6/61.5 |
MLVU | 48.3 | - | 68.2 |
LVBench | - | - | 43.3 |
MMBench - Video | 1.73 | 1.44 | 1.63 |
EgoSchema | - | - | 64.8 |
PerceptionTest | - | - | 66.9 |
TempCompass | - | - | 64.4 |
LongVideoBench | 55.2 | 55.6 | 54.2 |
CharadesSTA/mIoU | - | - | 38.8 |
Agent benchmark
Benchmarks | Qwen2.5 - VL - 3B |
---|---|
ScreenSpot | 55.5 |
ScreenSpot Pro | 23.9 |
AITZ_EM | 76.9 |
Android Control High_EM | 63.7 |
Android Control Low_EM | 22.2 |
AndroidWorld_SR | 90.8 |
MobileMiniWob++_SR | 67.9 |
Requirements
The code of Qwen2.5 - VL has been included in the latest Hugging face transformers. We advise you to build from source with the command:
pip install git+https://github.com/huggingface/transformers accelerate
Otherwise, you might encounter the following error:
KeyError: 'qwen2_5_vl'
Quickstart
Below, we provide simple examples to show how to use Qwen2.5 - VL with ๐ค ModelScope and ๐ค Transformers. The code of Qwen2.5 - VL has been included in the latest Hugging face transformers. We advise you to build from source with the command:
pip install git+https://github.com/huggingface/transformers accelerate
Otherwise, you might encounter the following error:
KeyError: 'qwen2_5_vl'
We offer a toolkit to help you handle various types of visual input more conveniently, as if you were using an API. This includes base64, URLs, and interleaved images and videos. You can install it using the following command:
# It's highly recommended to use `[decord]` feature for faster video loading.
pip install qwen - vl - utils[decord]==0.0.8
If you are not using Linux, you might not be able to install decord
from PyPI. In that case, you can use pip install qwen - vl - utils
, which will fall back to using torchvision for video processing. However, you can still [install decord from source](https://github.com/dmlc/decord?tab=readme - ov - file#install - from - source) to use decord when loading videos.







