Qwen3-VL-2B-GRACE-W8G128

This repository contains a GRACE-trained Qwen3-VL-2B checkpoint using quantization-aware training (QAT) with W8G128 group-wise INT8 quantization.

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
• Method: GRACE
• Quantization: W8G128 group-wise INT8 QAT
• Training data: ShareGPT4V
• Training / evaluation protocol: LLaVA-style multimodal evaluation
• Library: Hugging Face Transformers
• Repository: ForeverBlue/Qwen3-VL-2B-GRACE-W8G128

---

📊 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.

Model	Params	Precision	HallB	MMBench	ScienceQA	AI2D	MMMU	SEED	MMStar	Avg
Qwen3-VL-8B (teacher, ref.)	8B	BF16	61.1	84.5	85.0	85.7	69.6	77.5	70.9	76.3
Qwen3-VL-2B (baseline)	2B	BF16	51.4	78.4	81.4	76.9	53.4	71.2	58.3	67.3
Qwen3-VL-2B-GRACE	2B	BF16	66.9	86.4	86.2	81.3	72.1	76.7	67.3	76.7
Qwen3-VL-2B-GRACE (W8G128)	2B	INT8	66.1	85.5	85.3	80.4	71.3	75.9	66.5	75.9
Qwen3-VL-2B-GRACE (W4G128)	2B	INT4	65.4	84.6	84.3	79.5	70.5	75.1	65.8	75.0

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 W8G128 INT8 model retains 99% of the BF16 average.

---

🤗 Model Zoo

Model	Backbone	Bits	Group	Checkpoint description	HF Hub
Qwen3-VL-2B-GRACE-BF16	Qwen3-VL-2B	bf16	—	Full-precision GRACE checkpoint; used as the student initialization for the W8/W4 Qwen3-VL runs.	FoeverBLUE/Qwen3-VL-2B-GRACE-BF16
Qwen3-VL-2B-GRACE-W8G128	Qwen3-VL-2B	int8	128	INT8 QAT checkpoint with group size 128; high-retention quantized Qwen3-VL student.	FoeverBLUE/Qwen3-VL-2B-GRACE-W8G128
Qwen3-VL-2B-GRACE-W4G128	Qwen3-VL-2B	int4	128	INT4 QAT checkpoint with group size 128; compact Qwen3-VL release retaining about 98% of the BF16 average.	FoeverBLUE/Qwen3-VL-2B-GRACE-W4G128
LLaVA-1.5-7B-GRACE-W4G128	LLaVA-1.5-7B	int4	128	INT4 QAT checkpoint from the GRACE paper with learned scales; released for reproducing the LLaVA-1.5 experiments.	FoeverBLUE/LLaVA-1.5-7B-GRACE-W4G128

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 purposes, including:

• Efficient vision-language models
• Quantization-aware training
• Low-bit multimodal model deployment
• Knowledge distillation for VLM compression
• Multimodal model efficiency studies

---

Out-of-Scope Use

This checkpoint is not intended for:

• Safety-critical deployment
• Medical / legal / financial decision-making
• Production systems requiring reliability guarantees

Like other VLMs, the model may generate hallucinated, biased, or incorrect outputs.

---

Training Data

The model was trained using ShareGPT4V multimodal instruction data under a LLaVA-style multimodal fine-tuning pipeline.

Dataset:

• Lin-Chen/ShareGPT4V

---

Quantization Details

This checkpoint uses quantization-aware training (QAT) with group-wise W8G128 quantization.

Configuration:

• Weight precision: INT8
• Group size: 128
• Quantization scheme: Group-wise QAT
• Method: GRACE
• Backbone: Qwen3-VL-2B-Instruct

Depending on the inference backend, specialized quantized kernels or custom loading logic may be required to obtain real INT8 deployment benefits.

---

Repository Files

This repository may contain:

• model.safetensors / model-*.safetensors — model weights
• qat_quantized_weights.bin — QAT quantized weight artifact
• config.json — model configuration
• generation_config.json — generation configuration
• tokenizer files
• processor / preprocessing configuration files

---

Loading

Please use a Qwen3-VL-compatible Transformers environment or the official Qwen3-VL codebase.

from transformers import AutoProcessor
from transformers import AutoModelForImageTextToText

repo_id = "ForeverBlue/Qwen3-VL-2B-GRACE-W8G128"

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

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

Recommended:

• recent transformers version
• Qwen3-VL compatible environment
• CUDA GPU inference backend for large-scale evaluation

---

Evaluation

The checkpoint follows a LLaVA-style multimodal evaluation protocol.

Representative evaluation may include benchmarks such as:

• HallusionBench
• MMBench
• ScienceQA
• AI2D
• MMMU
• SEED-Bench
• MMStar

Please refer to the associated GRACE paper and the results table above for detailed evaluation settings and results.

---

Important Notes

This checkpoint includes QAT-specific quantized weights in qat_quantized_weights.bin. Depending on the inference codebase, additional GRACE-specific quantization-aware loading logic may be required.

The standard from_pretrained call may load the model configuration and checkpoint files, but fully reproducing the intended INT8 QAT behavior may require the GRACE repository:

https://github.com/ForeverBlue816/GRACE

---

Limitations

• This model is released for research purposes.
• The quantized checkpoint may require custom loading logic for QAT-specific weights.
• 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.
• Specialized kernels or custom loading code may be required to realize practical INT8 speed or memory benefits.

---

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}
}

Please also cite the original Qwen3-VL work when using this model.

---

License

Released under the MIT license.

Users should additionally comply with:

• Qwen3-VL base model license
• ShareGPT4V dataset terms
• applicable downstream usage restrictions