Bowen(Brian) Qu - Homepage

Bowen Qu Brian

About Me

👋Hi, I’m Bowen(Brian) Qu, doing PostTrain at Moonshot.ai (Kimi, 月之暗面). My research interests include MLLM, TIR(Tool-Integrated Reasoning) and Agentic RL. The logo of this website is my lovely cat - Baka (巴卡 in Chinese)!

Education: I received my MPhil. degree at School of Electronic and Computer Science (SECE), Peking University (PKU), in June, 2025. Previously, I received the Honours B.E. degree at School of Electronic Information and Communications (EIC), Huazhong University of Science and Technology (HUST) in June, 2022.
Experience: Fortunately, I have the honor to participate in some interesting MLLM research projects:
- [2025.03 - Present] Moonshot.ai, Technical Staff (PostTrain). Cooking K2.5 series:
  - 🔹 Native Multimodal RL (K2.5 Report Chap2.2, ZeroVision ColdStart -> Vision-Centric RL)
  - 🔹 Vision Reasoning & Knowledge & TIR RL
  - 🔹 Chart Understanding and Chart-to-Code
- [2024.04 - 2024.12] 01.ai & Rhymes.ai, Research Intern, Multimodal Team, supervised by Junnan Li, working closely with Dongxu Li and Haoning Wu.
  - 🏆 Core Contributor of Aria — an Open Multimodal Native MoE
- [2024.02 - 2024.07] IDEA Research, Research Intern, working closely with Zhengzhuo Xu, Yiyan Qi and Chengjin Xu.
  - 🏆 Co-first Author of ChartMoE (ICLR2025 Oral): Mixture of Diversely Aligned Expert Connector for Chart Understanding
Status: I'm always eager to learn new insights and ideas. The potential of Agents are still under exploration. If you're in Beijing, let's grab coffee to discuss it. Also feel free to drop me an 📧 if there is a good fit!

Interests

Vision-Language Model(VLM)
MLLM Reasoning
Al-Generated Image/Video Quality Assesment

Education

Master of Science

Peking University (PKU)
Bachelor of Engineering

Huazhong University of Science and Technology (HUST)

🔥 News

2026.01: 💥💥💥 We release Kimi-K2.5❗️❗️❗️
2025.11: 💥 We release IE-Critic-R1, a MLLM specialized in assessing the quality of text-driven image editing results. It is a Pointwise, Generative Reward Model, leveraging CoT reasoning SFT and RLVR to provide accurate, human-aligned evaluations of image editing.
2025.07: 💥💥💥 We release Kimi-K2❗️❗️❗️
2025.02: 🎉🎉🎉 ChartMoE is selected as ICLR2025 Oral(1.8%)!
2025.01: 🎉🎉 ChartMoE is accepted by ICLR2025!
2024.10: 🎉🎉 We release Aria, a native LMM that excels on text, code, image, video, PDF and more!
2024.09: 💥💥 We release ChartMoE, a MLLM with MoE connector, for advanced chart 1️⃣understanding, 2️⃣replot, 3️⃣editing, 4️⃣highlighting and 5️⃣transformation.
2023.12: 💥 MPP-Qwen-Next is released! Prevent poverty (24GB of VRAM) from limiting imagination. All 7B/14B llava-like training is conducted on RTX3090 GPUs by Pipeline Parallel.

Selected Outputs

🌟 is me. * Equal Contribution (i.e.: Co-First Author). 📧 Corresponding Author.

Native Multimodal RL

Kimi-K2.5: Visual Agentic Intelligence

Kimi-K2.5 is Moonshot's most powerful native multimodal agentic model (1T params, 32B activated). I led the Native Multimodal RL effort (K2.5 Report Chap2.2): ZeroVision ColdStart → Vision-Centric RL, plus Vision Reasoning & Knowledge & TIR RL.

Mar 27, 2025

MoE in Downstream Tasks

ChartMoE: Mixture of Diversely Aligned Expert Connector for Chart Understanding

ChartMoE is a multimodal large language model with Mixture-of-Expert connector for advanced chart 1️⃣understanding, 2️⃣replot, 3️⃣editing, 4️⃣highlighting and 5️⃣transformation.

Jan 22, 2025

Multimodal Native MoE

Aria: An Open Multimodal Native Mixture-of-Experts Model

Aria is a multimodal native MoE model. It features：1️⃣State-of-the-art performance on various multimodal and language tasks, superior in video and document understanding; 2️⃣Long multimodal context window of 64K tokens; 3️⃣3.9B activated parameters per token, enabling fast inference speed and low fine-tuning cost.

Oct 5, 2024

Core-Authored Publications

🌟 is me. * Equal Contribution (i.e.: Co-First Author). 📧 Corresponding Author.

🌟Bowen Qu, Shangkun Sun, Xiaoyu Liang, Wei Gao📧 (2025). IE-Critic-R1: Advancing the Explanatory Measurement of Text-Driven Image Editing for Human Perception Alignment. arXiv preprint.

PDF Code Cite Model Dataset

Zhengzuo Xu*, 🌟Bowen Qu*, Yiyan Qi*, Sinan Du, Chengjin Xu, Chun Yuan📧, Jian Guo📧 (2025). ChartMoE: Mixture of Diversely Aligned Expert Connector for Chart Understanding. ICLR2025 Oral (1.8%).

PDF Cite Code Dataset Slides Source Document

Dongxu Li, Yudong Liu, Haoning Wu, Yue Wang, Zhiqi Shen, 🌟Bowen Qu, Xinyao Niu, Fan Zhou, Chengen Huang, Yanpeng Li, Chongyan Zhu, Xiaoyi Ren, Chao Li, Yifan Ye, Peng Liu, Lihuan Zhang, Hanshu Yan, Guoyin Wang, Bei Chen, Junnan Li📧 (2024). Aria: An Open Multimodal Native Mixture-of-Experts Model. Technical Report.

PDF Cite Code Dataset Slides Source Document

🌟Bowen Qu, Xiaoyu Liang, Shangkun Sun, Wei Gao📧 (2024). Exploring AIGC Video Quality: A Focus on Visual Harmony, Video-Text Consistency and Domain Distribution Gap. CVPRW2024.

PDF Cite Code Poster

🌟Bowen Qu*, Haohui Li*, Wei Gao📧 (2024). Bringing Textual Prompt to AI-Generated Image Quality Assessment. ICME2024.

PDF Cite Code Poster

Selected Projects

😺 I enjoy open-sourcing. Here are a selection of projects that I’ve led or served as the core contributor.

Native Multimodal RL