what's the base model?
Thanks for sharing! is the base model a qwen-3 32B? or it's a more specialized model on deep research?
We then ran CPT, SFT, and RL to specialize it for deep-research tasks.
We tested it on our own dataset of 1,000 prompts focusing on action interpretation and sequencing. This is not a public benchmark.
Under the same parameter settings, the performance is Qwen3-Next-80B-A3B > Qwen3-30B-A3B > Tongyi-DeepResearch-30B-A3B.
vLLM offline inference without quantization: temperature=0.6, top_p=0.95, top_k=20, repetition_penalty=1.0, presence_penalty=0.0, seed=42
Is this performance trend common across other tasks?
The base model is Qwen3-30B-A3B-Thinking-2507.
The paper https://arxiv.org/pdf/2509.13310 says otherwise:
"Starting from Qwen’s pre-trained foundation models (e.g., Qwen3-30B-A3B-Base), our enhanced training pipeline consists of"
So just wanted to confirm which one is it? Thanks for the help!
Thank you for your attention. For AgentFounder and Tongyi Deep Research, our model's starting point is Qwen3-30B-A3B-Base, upon which we conduct Agentic Continual pre-training. In addition, our other post-training work may be based on Qwen3-30B-A3B-Thinking-2507, see the description in the paper for details.