Qwen3-VL-2B-GRACE-BF16

This repository contains a full-precision BF16 checkpoint of Qwen3-VL-2B trained using the GRACE framework.

Start here: Upload an image and try the GRACE 2B student in the GRACE-VLM Space. For genuine packed INT4 deployment, use the primary Qwen3-VL-2B-GRACE-W4G128-AWQ checkpoint and its copy-ready GRACE loader command.

This model is associated with our ICML 2026 paper:

Gated Relational Alignment via Confidence-based Distillation for Efficient VLMs
Yanlong Chen, Amirhossein Habibian, Luca Benini, Yawei Li
Accepted to the International Conference on Machine Learning (ICML 2026)

• Paper: https://arxiv.org/abs/2601.22709
• DOI: https://doi.org/10.48550/arXiv.2601.22709
• Code: https://github.com/ForeverBlue816/GRACE

---

Model Details

• Base model: Qwen/Qwen3-VL-2B-Instruct
• Training framework: GRACE
• Precision: BF16 full precision
• Training data: ShareGPT4V
• Evaluation protocol: LLaVA-style multimodal evaluation
• Repository: ForeverBlue/Qwen3-VL-2B-GRACE-BF16

---

📊 Results

Comparison on 7 VLM benchmarks. The 8B model is the distillation teacher (reference upper bound); all GRACE-Qwen3 variants are 2B students. Best result among the 2B Qwen3-VL models is in bold.

We release GRACE on Qwen3-VL here because it is the most current backbone and gives a fairer, up-to-date point of comparison, with the vanilla Qwen3-VL-2B-Instruct as the baseline. The paper itself reports GRACE on LLaVA-1.5 and Qwen2-VL; we additionally release the LLaVA-1.5 W4G128 INT4 checkpoint from the paper in the model zoo below.

GRACE lifts the Qwen3-VL-2B baseline by +9.4 avg and matches or slightly exceeds the 8B teacher on average (76.7 vs. 76.3) at roughly 1/4 the parameters. The W4G128 INT4 model retains 98% of the BF16 average.

---

🤗 Model Zoo

The BF16 Qwen3-VL checkpoint is the full-precision GRACE student used as the initial student weights for the W8 and W4 Qwen3-VL runs. The LLaVA-1.5 W4G128 checkpoint corresponds to the paper setting and includes GRACE-specific QAT quantized weights for reproducing the INT4 LLaVA experiments.

---

Intended Use

This model is intended for research on:

• Efficient vision-language models
• Knowledge distillation for VLMs
• Multimodal alignment
• Full-precision GRACE training
• BF16 baseline / teacher-student comparison studies

---

Training Details

This checkpoint is a full-precision BF16 model trained under the GRACE framework.

Configuration:

• Precision: BF16
• Training method: GRACE
• Backbone: Qwen3-VL-2B-Instruct
• Dataset: ShareGPT4V
• Evaluation: LLaVA-style multimodal benchmarks

Unlike the QAT releases, this model does not use weight quantization.

---

Files

• model.safetensors / model-*.safetensors
• config.json
• generation_config.json
• tokenizer files
• processor files

---

Loading

from transformers import AutoProcessor
from transformers import AutoModelForImageTextToText
import torch

repo_id = "ForeverBlue/Qwen3-VL-2B-GRACE-BF16"

processor = AutoProcessor.from_pretrained(
    repo_id,
    trust_remote_code=True
)

model = AutoModelForImageTextToText.from_pretrained(
    repo_id,
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

---

Important Notes

This is the full-precision BF16 GRACE checkpoint. It does not include INT8 or INT4 QAT weight compression. For quantized versions, please refer to the W8G128 and W4G128 checkpoints listed in the Model Zoo.

The standard from_pretrained call should load this BF16 checkpoint directly in a Qwen3-VL-compatible Transformers environment. For reproducing the GRACE training or distillation pipeline, please refer to the official code repository:

https://github.com/ForeverBlue816/GRACE

---

Limitations

• This model is released for research purposes.
• Performance may vary depending on the evaluation codebase, preprocessing, generation parameters, and multimodal benchmark implementation.
• Users should follow the license and usage restrictions of the original Qwen3-VL-2B-Instruct base model.
• This checkpoint is not optimized for low-bit inference; use the W8G128 or W4G128 release for quantized deployment studies.

---

Citation

If you use this model, please cite:

@article{chen2026gated,
  title={Gated Relational Alignment via Confidence-based Distillation for Efficient VLMs},
  author={Chen, Yanlong and Habibian, Amirhossein and Benini, Luca and Li, Yawei},
  journal={arXiv preprint arXiv:2601.22709},
  year={2026}
}