Kaixin Li

Final year PhD @ NUS
Prev Qwen, Tiktok, Sea AI Lab

Agents, Code and Multimodality

Core contributor of Qwen3.5 & VL &Coder

🖥️ScreenSpot-Pro

[Kaixin Li @kxli_2000]

@tbpn @JustinLin610 @huybery Stop it, I was just an intern. Junyang and Binyuan were the real cores and leads of the team.

[Tianyi Bai @becca_bai]

Junyang has been the most supportive mentor since I started my internship. When I took over the computer use agent project, we built the RL infrastructure from scratch. From the massive engineering challenges to resource coordination, he made it possible.

[Kaixin Li @kxli_2000]

@kyz18300422 @Bandana_ici @Yulun_Du @yzhang_cs real-real

[Kaixin Li @kxli_2000]

@Alibaba_Qwen Qwen could have had a Singapore base, all thanks to Junyang. But now that he's gone, there's no reason left to stay here.

[Kaixin Li @kxli_2000]

Signing off from @Alibaba_Qwen. Grateful for the chance to work with such brilliant minds. Proud of our impact. Onwards and upwards!

[Yang Su @YangSu2000]

@JustinLin610 qwen is nothing without its people...

[Kaixin Li @kxli_2000]

@JustinLin610 Best boss ever! Thank you for everything!

[Qwen @Alibaba_Qwen]

🚀 Introducing the Qwen 3.5 Small Model Series
Qwen3.5-0.8B · Qwen3.5-2B · Qwen3.5-4B · Qwen3.5-9B

✨ More intelligence, less compute.
These small models are built on the same Qwen3.5 foundation — native multimodal, improved architecture, scaled RL:
• 0.8B / 2B → tiny, fast, great for edge device
• 4B → a surprisingly strong multimodal base for lightweight agents
• 9B → compact, but already closing the gap with much larger models
And yes — we’re also releasing the Base models as well.
We hope this better supports research, experimentation, and real-world industrial innovation.
Hugging Face: https://t.co/wFMdX5pDjU
ModelScope: https://t.co/9NGXcIdCWI

[Kaixin Li @kxli_2000]

The task is from MCP-Universe github_task_0011.

Now imagine a human spending hours debugging, unknowingly training a lazy bot that outsourced its own benchmark failure to you.

[Kaixin Li @kxli_2000]

We used to fear agents taking our jobs.
🤖🫵 Instead, witness the first case of Human-as-a-Tool.

Model: "I can't finish this MCP task."
Also Model: Opens GitHub Issue "Hey human, can you fork this repo and do all the work for me?"

We are now working for the agents.

[Junyang Lin @JustinLin610]

Qwen3.5 is Live! Today we openweight the first model, Qwen3-397B-A17B, which is a native multimodal model supporting both thinking and non-thinking modes. We have strengthened its coding and agentic capabilities to foster productivity for developers and enterprises. Hope you enjoy it, and more open-weight models are coming in this Chinese New Year!

[Kaixin Li @kxli_2000]

Just shipped updates across Visual Programming, WebDev, GUI, and MCP in this model! 🚀

Would love for you to test it out and share your thoughts.

[Qwen @Alibaba_Qwen]

🚀 Qwen3.5-397B-A17B is here: The first open-weight model in the Qwen3.5 series.

🖼️Native multimodal. Trained for real-world agents.
✨Powered by hybrid linear attention + sparse MoE and large-scale RL environment scaling.
⚡8.6x–19.0x decoding throughput vs Qwen3-Max
🌍201 languages & dialects
📜Apache2.0 licensed

🔗Dive in:
GitHub: https://t.co/NzNdS9joAT
Chat: https://t.co/bg4tAU0Rhw
API：https://t.co/YiiyKTnHoU
Qwen Code: https://t.co/qqwj5nAger
Hugging Face: https://t.co/wFMdX5p5um
ModelScope: https://t.co/9NGXcId57a
blog: https://t.co/AW8UQStXaL

[Qwen @Alibaba_Qwen]

🚀 Introducing Qwen3-Coder-Next, an open-weight LM built for coding agents & local development.
What’s new:
🤖 Scaling agentic training: 800K verifiable tasks + executable envs
📈 Efficiency–Performance Tradeoff: achieves strong results on SWE-Bench Pro with 80B total params and 3B active
✨ Supports OpenClaw, Qwen Code, Claude Code, web dev, browser use, Cline, etc
🤗 Hugging Face: https://t.co/rZoW4vRJpr
🤖 ModelScope: https://t.co/P0vT5zILBZ
📝 Blog: https://t.co/hFfFDYcwvd
📄 Tech report: https://t.co/Qx83PWS3oi

[Qwen @Alibaba_Qwen]

🚀 Qwen3-VL Tech report is now out on arXiv!

From pretraining to post-training, architecture to infra, data to evaluation — we’ve packed in the details for anyone building on vision-language models.

🔥 3 models >1M downloads in just over a month
🏆 Qwen3-VL-8B leads with 2M+ downloads
📚 Built on the shoulders of Qwen2.5-VL (2800+ citations in <10 months!)

Check out the paper for insights, baselines, and future directions.
Let’s keep pushing VLMs forward — together.
https://t.co/9D4gFoopcN

[Kevin Lin @KevinQHLin]

🤗Excited to open-source GroundCUA!

🚀A large-scale, human-annotated dataset for precise UI grounding to advance Computer-Use Agents.

- 3.56M+ high-quality human annotations
- 56K screenshots
- 87 desktop applications
- all datasets and models are available

project page: https://t.co/X6UuITG5TK
arxiv: https://t.co/3KX2qQ9htq
dataset: https://t.co/uNqBqG1jZM
model: https://t.co/1PQsK8kpEV

[Qwen @Alibaba_Qwen]

Introducing the compact, dense versions of Qwen3-VL — now available in 4B and 8B pairs, each with both Instruct and Thinking variants.

✅ Lower VRAM usage
✅ Full Qwen3-VL capabilities retained
✅ Strong performance across the board

Despite their size, they outperform models like Gemini 2.5 Flash Lite and GPT-5 Nano, and often beat them on benchmarks spanning STEM, VQA, OCR, video understanding, agent tasks, and more. In many cases, they even rival our flagship Qwen2.5-VL-72B from just six months ago!

Plus, FP8 versions are also available for efficient deployment.

Hugging Face: https://t.co/MdtyH8lhWL
ModelScope: https://t.co/LWwKz7VzJ8
Qwen3-VL-8B-Instruct API： https://t.co/DpJXAEDisI
Qwen3-VL-8B-Thinking API： https://t.co/7xFK5aqxZm
Cookbooks: https://t.co/T06q60PP62

[Junyang Lin @JustinLin610]

some models next week

[Terry Yue Zhuo @terryyuezhuo]

It’s so much fun working with the other 39 community members on this project!

Start to try out various frontier models in BigCodeArena today.

[Arena.ai @arena]

🏆 Qwen3-VL-235B-a22b-Instruct now the #1 Open Model in Text!

The thinking variant landed overall at #30.
https://t.co/43uToJ8XJw

[Junyang Lin @JustinLin610]

Super interesting! For Qwen3-VL, the performance is out of our expectations (and for Qwen3-Max it is out of our expectations as well lol).

Anyway, you can now use VL as an LLM for sure, which means that the next generation models might be VL by default.