---
library_name: transformers
license: apache-2.0
base_model: Qwen/Qwen3.5-35B-A3B
base_model_relation: finetune
pipeline_tag: image-text-to-text
language:
- en
- zh
tags:
- exomind
- scientific-reasoning
- scientific-research
- agentic
- tool-use
- multimodal
- vision-language
- qwen3.5
- safetensors
---
## ๐ฅ News
- **2026-08-12**: ๐ฅ We release the ExoMind technical report, official project page,
and public repository.
## Overview
ExoMind is the first extended-mind-inspired agentic system designed for
scientific reasoning and research. It organizes a general-purpose model,
specialized interaction objects, and autonomous interaction processes as one
system, allowing the model to discover sources, ground evidence, execute
verification, and update its reasoning around each scientific problem.
This repository hosts the main checkpoint, fine-tuned from
[Qwen3.5-35B-A3B](https://huggingface.co/Qwen/Qwen3.5-35B-A3B). With
training-value-aware data engineering, a scientific interaction framework, and
two-stage progressive Chain-of-Interaction training, ExoMind raises the average
score across eight scientific benchmarks from **36.2 to 67.5**, achieves the
highest average among all evaluated models, and ranks first on six benchmarks.
## Highlights
- **Extended-mind-inspired intelligence:** unifies the LLM, interaction
objects, and autonomous interaction processes as a scientific agentic system.
- **Training-value-aware data engineering:** identifies challenging, learnable
problems and routes them to pure-reasoning or interaction-reasoning data.
- **Scientific interaction:** turns source discovery, evidence grounding,
executable verification, and observation integration into composable objects.
- **Progressive CoI training:** jointly develops intrinsic reasoning and
autonomous interaction using a few thousand high-quality trajectories.
- **Efficient frontier performance:** completes two-stage full-parameter SFT in
1โ2 days on 8 NVIDIA H200 GPUs while improving all six evaluated general
capability benchmarks over the base model.
## Quick Start
Use a recent vLLM or SGLang release with Qwen3.5 support. The examples below
use the checkpoint's configured maximum context length of 262,144 tokens.
### vLLM
```bash
vllm serve AI4SGI/ExoMind \
--port 8000 \
--tensor-parallel-size 8 \
--max-model-len 262144 \
--reasoning-parser qwen3 \
--enable-auto-tool-choice \
--tool-call-parser qwen3_coder
```
### SGLang
```bash
python -m sglang.launch_server \
--model-path AI4SGI/ExoMind \
--host 0.0.0.0 \
--port 8000 \
--tp-size 8 \
--context-length 262144 \
--reasoning-parser qwen3 \
--tool-call-parser qwen3_coder
```
### OpenAI-Compatible API
```python
from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")
response = client.chat.completions.create(
model="AI4SGI/ExoMind",
messages=[
{
"role": "user",
"content": "Develop and verify a rigorous solution to this scientific problem: ...",
}
],
temperature=1.0,
top_p=0.95,
extra_body={"top_k": 20},
)
print(response.choices[0].message.content)
```
The complete scientific interaction workflow and prompt contracts are available
in the [ExoMind GitHub repository](https://github.com/AI4SGI/ExoMind).
## Evaluation
Under the technical report's evaluation setup, ExoMind reaches an
eight-benchmark average of **67.5**, compared with **54.2** for the next-best
representative model shown below.
๐ฅ Best score among the representative models shown
| Benchmark |
โญ Ours |
Representative frontier models |
ExoMind
35B-A3B |
Claude-Opus-4.8
Thinking |
GPT-5.5
(xhigh) |
Gemini-3.1-Pro
Preview |
Kimi-K3 |
Qwen3.7-Max |
GLM-5.2 |
DeepSeek-V4-Pro
(Max) |
| ๐งช Scientific Research |
| HLE w/ tools | 50.9 | ๐ฅ 57.9 | 52.2 | 51.4 | 56.0 | 53.5 | 54.7 | 48.2 |
| FrontierScience-Research | ๐ฅ 70.0 | 26.7 | 26.7 | 11.7 | 21.7 | 10.0 | 15.0 | 13.3 |
| CMT-Benchmark | ๐ฅ 84.0 | 46.0 | 43.0 | 43.0 | 34.0 | 34.0 | 20.0 | 28.0 |
| CritPt | 25.7 | 20.9 | ๐ฅ 27.1 | 17.7 | 23.4 | 13.4 | 20.9 | 7.1 |
| ๐ง Scientific Reasoning |
| AMO-Bench | ๐ฅ 78.0 | 74.0 | 70.0 | 63.1 | 64.0 | 57.4 | 54.0 | 68.0 |
| IMO-AnswerBench | ๐ฅ 92.8 | 86.8 | 83.8 | 90.0 | 82.8 | 90.0 | 91.0 | 89.8 |
| HiPhO | ๐ฅ 49.7 | 46.4 | 43.3 | 43.4 | 42.4 | 38.8 | 37.4 | 38.7 |
| FrontierScience-Olympiad | ๐ฅ 89.0 | 75.0 | 78.0 | 77.0 | 69.0 | 80.0 | 76.5 | 76.0 |
| Eight-benchmark average | ๐ฅ 67.5 | 54.2 | 53.0 | 49.7 | 49.2 | 47.1 | 46.2 | 46.1 |
See the [interactive evaluation
explorer](https://ai4sgi.github.io/ExoMind/#results) for the complete model list,
benchmark scopes, settings, and rankings.
## Intended Use
ExoMind is intended for research and development in scientific question
answering, literature-grounded investigation, mathematical and computational
reasoning, code-assisted verification, and agentic scientific workflows.
## License and Attribution
The distributed checkpoint and upstream Qwen3.5 materials are provided under
the Apache License 2.0 included in this repository. The technical report,
scientific figures and results, and ExoMind brand assets are subject to the
[ExoMind Research Content and Brand Terms](./CONTENT_RIGHTS.md). See
[NOTICE.md](./NOTICE.md) for third-party notices.
## Citation
```bibtex
@misc{exomind2026,
title = {ExoMind: Democratizing Scientific Intelligence via Extended-Mind-Inspired Agentic System},
author = {Peng Ye and Zhuo Liu and Jingqi Ye and Fangchen Yu and Shengji Tang and Yichen Jiang and Haonan He and Zongsheng Cao and Tao Chen and Bo Zhang and Wanli Ouyang and Bowen Zhou and Lei Bai},
year = {2026},
note = {Technical report},
url = {https://github.com/AI4SGI/ExoMind/blob/main/Paper.pdf}
}
```
# ExoMind: Democratizing Scientific Intelligence via Extended-Mind-Inspired Agentic System
**ExoMind Team ยท Shanghai Artificial Intelligence Laboratory**
