We are recruiting high-level interns (Qingyun/้ไบ) for exploring scaling laws in agentic mid-training. If you are interested, feel free to contact me.
Junrulu
Junrulu
AI & ML interests
None yet
Recent Activity
authored a paper 1 day ago
Code-as-Room: Generating 3D Rooms from Top-Down View Images via Agentic Code Synthesis upvoted a paper about 2 months ago
HISA: Efficient Hierarchical Indexing for Fine-Grained Sparse AttentionOrganizations
replied to their post 4 months ago
posted an update 5 months ago
Post
281
We are pleased to introduce a brand-new lightweight LLM, Youtu-LLM:
(1) Youtu-LLM has a total of 2B parameters, employing a 32-layer dense MLA architecture and equipped with an innovative STEM- and agentic-oriented vocabulary;
(2) Based on approximately 11T tokens of pre-training, particularly with native 128k long context extension and native agentic mid-training, Youtu-LLM-2B is comparable to Qwen3-4B in general and agent capabilities;
(3) We have open-sourced the Base/Instruct versions, as well as the evaluation code for reproducing the test metrics. In the technical report, we share our experience of native agentic pre-training in detail.
Youtu-LLM-2B is very suitable as a starting point for exploring on-device agent practice. Meanwhile, we are currently extending this paradigm to larger-scale explorations. We welcome more discussion and collaboration!
๐ Check the project here: https://github.com/TencentCloudADP/youtu-tip/tree/master/youtu-llm
๐ค Check the models here: https://huggingface.co/collections/tencent/youtu
(1) Youtu-LLM has a total of 2B parameters, employing a 32-layer dense MLA architecture and equipped with an innovative STEM- and agentic-oriented vocabulary;
(2) Based on approximately 11T tokens of pre-training, particularly with native 128k long context extension and native agentic mid-training, Youtu-LLM-2B is comparable to Qwen3-4B in general and agent capabilities;
(3) We have open-sourced the Base/Instruct versions, as well as the evaluation code for reproducing the test metrics. In the technical report, we share our experience of native agentic pre-training in detail.
Youtu-LLM-2B is very suitable as a starting point for exploring on-device agent practice. Meanwhile, we are currently extending this paradigm to larger-scale explorations. We welcome more discussion and collaboration!
๐ Check the project here: https://github.com/TencentCloudADP/youtu-tip/tree/master/youtu-llm
๐ค Check the models here: https://huggingface.co/collections/tencent/youtu
- 1 reply