github.com/OceanGPT/OceanGPT @main

    English</a>&nbsp ｜ &nbsp<a href="https://github.com/OceanGPT/OceanGPT/raw/main/README_CN.md">中文</a>

沧渊海洋基础大模型：Ocean Foundation Models

Project • Paper • Models • Web • Manual • Overview • Quickstart • Citation

---

✨ OceanGPT Beginner's Guide |新手教程中文版officially published!

✨ OceanGPT Fine-tuning's Guide |定制问答引擎教程中文版officially published!

We have published a detailed beginner's guide for OceanGPT to help you quickly understand its capabilities. If you're looking to customize OceanGPT for practical use, you can refer to the Fine-tuning Guide to build a tailored question-answering engine.

[!IMPORTANT] We release OceanPile to provide a large-scale multimodal corpus for foundation models. We regularly update our open-source models, so their capabilities may differ from previous versions. We warmly welcome your feedback to help us continuously improve the application of LLMs in the ocean domain.

Table of Contents

• What's New
• Overview
• Quickstart
• 🤗Chat with Our Demo on Gradio
• Fine-tuning OceanGPT for Building Customized QA Applications
• Using MCP Server for Sonar Image Caption
• Inference
  • Models
  • Efficient Inference with sglang, vLLM, ollama, llama.cpp
• Acknowledgement
• Limitations
• Citation

🔔News

• 2025-12-15, we release new instruction-tuned and reasoning models, including OceanGPT-30B and OceanGPT-4B, using Huawei Ascend AI 910B.
• 2025-10-1, we released the initial version of OceanGym along with the accompanying paper.
• 2025-08-05, we release a tutorial on fine-tuning the OceanGPT model for Task-Oriented QA tasks.
• 2025-06-17, we release the OceanGPT-coder-0.6B.
• 2025-05-29, we deploy the OceanGPT MCP Server to support sonar image interpretation.
• 2025-04-20, we release the OceanGPT-o-7B and OceanGPT-coder-7B.
• 2025-02-01, we collect sonar data for model training and test OceanGPT-coder.
• 2024-12-01, we collect more publicly available sonar data and scientific images for model training.
• 2024-08-01, we launch bilingual (Chinese-English) multimodal large language model OceanGPT-o with sonar and ocean science image data collection and training.
• 2024-07-04, we release the OceanGPT-basic-14B/2B and the updated version of OceanGPT-basic-7B (v0.2).
• 2024-06-04, OceanGPT is accepted by ACL 2024. 🎉🎉
• 2023-10-04, we release the paper "OceanGPT: A Large Language Model for Ocean Science Tasks" and release OceanGPT-basic-7B (v0.1) based on LLaMA2.
• 2023-05-01, we launch the OceanGPT (沧渊) project.

---

Models

Model Name	ModelScope	HuggingFace
OceanGPT-basic-30B-A3B-Instruct (based on Qwen, recommended)	30B-A3B	30B-A3B
OceanGPT-basic-30B-A3B-Thinking (based on Qwen, recommended)	30B-A3B	30B-A3B
OceanGPT-basic-4B-Instruct (based on Qwen, recommended)	4B	4B
OceanGPT-basic-4B-Thinking (based on Qwen, recommended)	4B	4B
OceanGPT-o-7B (based on Qwen, recommended)	7B	7B
OceanGPT-coder-7B (based on Qwen, recommended)	7B	7B
OceanGPT-basic-8B (based on Qwen)	8B	8B
OceanGPT-basic-14B (based on Qwen, legacy)	14B	14B
OceanGPT-basic-7B (based on Qwen, legacy)	7B	7B
OceanGPT-basic-2B (based on MiniCPM, legacy)	2B	2B
OceanGPT-coder-0.6B (based on Qwen3)	0.6B	0.6B

---

• ❗Please note that the ocean domain Q&A in the online demo system (including the video) is based on knowledge base augmentation and a "general-specialized integration" approach, and the generated content differs from that of the open-source models (注意：在线演示系统和视频里的海洋专业问答采用了知识增强与通专结合等技术，因此和开源模型存在差异)！
• ❗Due to limited computing resources, OceanGPT-o is currently only applicable for natural language interpretation and generation of certain types of sonar images and marine science images. It is recommended to use a GPU that is greater than or equal to 24GB.

Instruction Data

Data Name	HuggingFace	ModelScope
OceanInstruct-v0.2	50K	50K
OceanInstruct-o	50K	50K
OceanInstruct-v0.1	10K	10K
---
- ❗Some of the instruction data are synthetic data; we apologize for any inaccuracies that may exist (部分指令数据为合成数据，如存在错误敬请谅解)！

🌟Overview

This is the OceanGPT (沧渊) project, which aims to build ocean foundation model.

⏩Quickstart

conda create -n py3.11 python=3.11
conda activate py3.11
pip install -r requirements.txt

Download the model

Download from HuggingFace

# use git lfs
git lfs install
git clone https://huggingface.co/zjunlp/OceanGPT-o-7B

or

# use huggingface-cli
pip install -U huggingface_hub
huggingface-cli download --resume-download zjunlp/OceanGPT-o-7B --local-dir OceanGPT-o-7B --local-dir-use-symlinks False

Download from ModelScope

# use git lfs
git lfs install
git clone https://www.modelscope.cn/ZJUNLP/OceanGPT-o-7B.git

or

# use modelscope
pip install modelscope
modelscope download --model ZJUNLP/OceanGPT-o-7B

Inference

OceanGPT-basic-8B

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "zjunlp/OceanGPT-basic-8B"

# load the tokenizer and the model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

question = "<Your Question>"
messages = [
    {"role": "user", "content": question}
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False 
)

model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=8192
)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist() 

try:
    index = len(output_ids) - output_ids[::-1].index(151668)  # </think> token ID
except ValueError:
    index = 0

content = tokenizer.decode(output_ids[index:], skip_special_tokens=True).strip("\n")
print(content)

OceanGPT-o-7B

# It's highly recommanded to use `[decord]` feature for faster video loading.
pip install qwen-vl-utils[decord]==0.0.8
pip install transformers

from transformers import Qwen2_5_VLForConditionalGeneration, AutoTokenizer, Qwen2VLProcessor
from qwen_vl_utils import process_vision_info
import torch

model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    "zjunlp/OceanGPT-o-7B", torch_dtype=torch.bfloat16, device_map="auto"
)
processor = Qwen2VLProcessor.from_pretrained("zjunlp/OceanGPT-o-7B")

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "file:///path/to/your/image.jpg",
            },
            {"type": "text", "text": "Describe this image."},
        ],
    }
]

text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(
    text=[text],
    images=image_inputs,
    videos=video_inputs,
    padding=True,
    return_tensors="pt",
)
inputs = inputs.to("cuda")


generated_ids = model.generate(**inputs, max_new_tokens=128)
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)

OceanGPT-coder-7B

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "zjunlp/OceanGPT-coder-7B", torch_dtype=torch.float16, device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("zjunlp/OceanGPT-coder-7B")
messages = [
    {"role": "system", "content": "You are Qwen, created by Alibaba Cloud. You are a helpful assistant."},
    {"role": "user", "content": "请为水下机器人生成MOOS代码，实现如下任务：先回到（50,20）点，然后以（15,20）点为圆形，做半径为30的圆周运动，持续时间200s，速度4 m/s。"}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
generated_ids = model.generate(
    **model_inputs,
    top_p=0.6,
    temperature=0.6,
    max_new_tokens=2048
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]
response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(response)

Inference by vllm

from transformers import AutoTokenizer
from vllm import LLM, SamplingParams

path = 'YOUR-MODEL-PATH'

tokenizer = AutoTokenizer.from_pretrained(path)

prompt = "Which is the largest ocean in the world?"
messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)

sampling_params = SamplingParams(temperature=0.8, top_k=50)
llm = LLM(model=path)

response = llm.generate(text, sampling_params)

🤗Chat with Our Demo on Gradio

Local WebUI Demo

You can easily deploy the interactive interface locally using