Recon2Reason-Reasoning-4B

Recon2Reason Reasoning 4B#

Recon2Reason Reasoning 4B is a 4B-scale vision-language model specialized for spatial reasoning in indoor and embodied scenes. It is fine-tuned from Qwen3-VL-4B and improves reasoning about metric distance, relative position, object configuration, and spatial relations from visual inputs.

The checkpoint retains the standard Qwen3VLForConditionalGeneration architecture. No custom model code or trust_remote_code=True is required.

Highlights#

• Standard Qwen3-VL Transformers interface
• 4,437,815,808 parameters
• BF16 weights in sharded Safetensors format
• Single-image and multi-image visual inputs
• Strong results on metric and qualitative spatial reasoning benchmarks

This repository contains the reasoning model only. The retrieval-augmented scene-reconstruction extension is released separately.

Model details#

Quick start#

pip install "transformers==4.57.1" "torch>=2.6" accelerate safetensors pillow

import torch
from transformers import AutoProcessor, Qwen3VLForConditionalGeneration

model_id = "BAAI/Recon2Reason-Reasoning-4B"

processor = AutoProcessor.from_pretrained(model_id)
model = Qwen3VLForConditionalGeneration.from_pretrained(
    model_id,
    dtype="auto",
    device_map="auto",
).eval()

messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": "path/to/scene.jpg"},
            {
                "type": "text",
                "text": "Which object is closest to the chair? ",
            },
        ],
    }
]

inputs = processor.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_dict=True,
    return_tensors="pt",
).to(model.device)

with torch.inference_mode():
    output_ids = model.generate(
        **inputs,
        max_new_tokens=512,
        do_sample=False,
    )

generated_ids = output_ids[:, inputs["input_ids"].shape[1]:]
answer = processor.batch_decode(
    generated_ids,
    skip_special_tokens=True,
    clean_up_tokenization_spaces=False,
)[0]
print(answer)

The model uses the standard Qwen3-VL chat format. For benchmark reproduction, use greedy decoding (do_sample=False) unless a benchmark specifies otherwise.

Evaluation#

The checkpoint was evaluated in BF16 with PyTorch 2.8.0, Transformers 4.57.1, SDPA attention, and greedy decoding on NVIDIA RTX PRO 6000 Blackwell GPUs.

License#

The model is released under the Apache License 2.0, subject to final confirmation that all training data and upstream artifacts permit this distribution. See LICENSE.

Acknowledgements#

This work builds on Qwen3-VL. We thank the creators of the evaluation datasets and the open-source Transformers ecosystem.