Alibaba’s Tongyi Lab has launched Agentic Continual Pre-training, an open source framework that enables enterprises to build powerful research agents offline—cutting API costs and challenging proprietary systems.
Alibaba’s Tongyi Lab has announced Agentic Continual Pre-training (Agentic CPT), an open source training framework designed to create advanced large language model (LLM) research agents without incurring costly API calls. The framework introduces a pre-aligned agentic base model capable of strong planning and self-correction, addressing a critical performance gap between open-source and proprietary systems.
At the centre of this breakthrough is AgentFounder-30B, a research agent built on the open-source Qwen3-30B model. It outperformed all existing open-source competitors across four benchmarks, surpassing DeepSeek-V3.1 by 10 percentage points on English BrowseComp. It also became the first open source model to cross 30 points on the demanding Humanity’s Last Exam (HLE) benchmark and scored 75.3% on Academic Browse, proving its academic utility.
The cost advantage comes from Agentic CPT’s offline data synthesis pipeline, which uses First-order Action Synthesis (FAS) and Higher-order Action Synthesis (HAS) to generate large-scale training data without reliance on external APIs or human annotation.
“Other deep research agents such as WebSailor, only post-train the model on a small amount of trajectories. Agentic CPT adds a new stage before them, which introduces a large scale of agentic data for stronger ability on agentic tasks,” explained Xinyu Wang, Researcher at Alibaba and Co-author of the Paper.
For enterprises, the framework promises faster, more affordable custom agents deployable on-premise, with enhanced reliability and risk control. As Wang added, “With a strong pre-aligned agentic base model, enterprises can perform light adaptation using their in-domain corpora and proprietary tools to rapidly build domain-specific agents… This approach is feasible in both cost and timeline for most companies.”
With Agentic CPT, Alibaba positions open source as a viable alternative to closed research models, combining scalability, affordability, and enterprise control.
