Octopus-8B

Octopus-8B is built based on Qwen-3-VL-8B-Instruct, featuring self-correction reasoning ability.

Paper:

Project Page: https://dripnowhy.github.io/Octopus/

Code: https://github.com/DripNowhy/Octopus

This is the weight repository for Octopus-8B.

---

Model Performance

Quickstart

Below, we provide simple examples to show how to use $\texttt{Octopus-8B}$ with vLLM and 🤗 Transformers.

First, Qwen3-VL has been in the latest Hugging Face transformers and we advise you to build from source with command:

pip install git+https://github.com/huggingface/transformers
# pip install transformers==4.57.0 # currently, V4.57.0 is not released

Using vLLM to Chat

Here we show a code snippet to show how to use the chat model with vllm:

from vllm import LLM, SamplingParams
from transformers import AutoProcessor
from PIL import Image

prompt_suffix = """\n\nYou first think through your reasoning process as an internal monologue, enclosed within <think> </think> tags. Then, provide your final answer enclosed within \\boxed{}. If you believe the answer can be further enhanced, generate <self-correction> </self-correction> tags enclosed with no content, and regenerate a new reasoning process and a new answer from scratch after that. The new response should first think through your reasoning process as an internal monologue, enclosed within <think> </think> tags. Then, provide your final answer enclosed within \\boxed{}. All reasoning, answer steps must be included without omission."""

MODEL_PATH = "Tuwhy/Octopus-8B"

def main():
    # Initialize model
    llm = LLM(
        model=MODEL_PATH,
        tensor_parallel_size=1,
        gpu_memory_utilization=0.9,
        seed=1,
        max_model_len=8192 * 8,
        trust_remote_code=True
    )

    processor = AutoProcessor.from_pretrained(
        MODEL_PATH,
        max_pixels=1280*28*28,
        min_pixels=256*28*28
    )

    # Single case
    prompt = "The accuracy gap between the Octopus-8B and the Qwen3-8B-VL-Thinking model is?"
    image_path = "./head.png"

    sampling_params = SamplingParams(
        temperature=1.0,
        top_p=0.95,
        top_k=-1,
        max_tokens=8192*2
    )

    # Prepare messages
    messages = [
        {
            "role": "user",
            "content": [
                {"type": "image", "image": image_path},
                {"type": "text", "text": prompt + prompt_suffix}
            ]
        }
    ]

    text_prompt = processor.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True
    )

    # Load image
    image = Image.open(image_path).convert("RGB")

    # Prepare input
    inputs = {
        "prompt": text_prompt,
        "multi_modal_data": {
            "image": image
        }
    }

    # Generate
    outputs = llm.generate([inputs], sampling_params=sampling_params)

    # Print result
    generated_text = outputs[0].outputs[0].text

    print("Generated response:")
    print("=" * 50)
    print(generated_text)
    print("=" * 50)

if __name__ == '__main__':
    main()

Using 🤗 Transformers to Chat

Here we show a code snippet to show how to use the chat model with transformers:

from transformers import Qwen3VLForConditionalGeneration, AutoProcessor

prompt_suffix = """\n\nYou first think through your reasoning process as an internal monologue, enclosed within <think> </think> tags. Then, provide your final answer enclosed within \\boxed{}. If you believe the answer can be further enhanced, generate <self-correction> </self-correction> tags enclosed with no content, and regenerate a new reasoning process and a new answer from scratch after that. The new response should first think through your reasoning process as an internal monologue, enclosed within <think> </think> tags. Then, provide your final answer enclosed within \\boxed{}. All reasoning, answer steps must be included without omission."""

# default: Load the model on the available device(s)
model = Qwen3VLForConditionalGeneration.from_pretrained(
    "Tuwhy/Octopus-8B", dtype="auto", device_map="auto"
)

# We recommend enabling flash_attention_2 for better acceleration and memory saving, especially in multi-image and video scenarios.
# model = Qwen3VLForConditionalGeneration.from_pretrained(
#     "Qwen/Qwen3-VL-8B-Instruct",
#     dtype=torch.bfloat16,
#     attn_implementation="flash_attention_2",
#     device_map="auto",
# )

processor = AutoProcessor.from_pretrained("Tuwhy/Octopus-8B")

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "./head.png",
            },
            {"type": "text", "text": "The accuracy gap between the Octopus-8B and the Qwen3-8B-VL-Thinking model is?" + prompt_suffix},
        ],
    }
]

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

# Inference: Generation of the output
generated_ids = model.generate(**inputs, max_new_tokens=8192*2)
generated_ids_trimmed = [
    out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print(output_text)

Generation Hyperparameters

VL

export greedy='false'
export top_p=0.95
export top_k=-1
export temperature=0.6
export out_seq_length=16384

Citation

If you find our work helpful, feel free to give us a cite.

@article{ding2025sherlock,
  title={Sherlock: Self-Correcting Reasoning in Vision-Language Models},
  author={Ding, Yi and Zhang, Ruqi},
  journal={arXiv preprint arXiv:2505.22651},
  year={2025}
}