According to TrendForce's Beating Monitor, Chen Tianqiao's new AI project, Apodex, has released the deep-research intelligent agent system, Apodex 1.0, focusing on achieving verifiable automation research through asynchronous collaboration and an evidence graph. Unlike traditional query-based large language models, Apodex 1.0 is designed for long-term research, incorporating planning, tool invocation, hypothesis testing, and evidence synthesis mechanisms.
In the Reloaded Deployment mode (Apodex-1.0-H), the system is orchestrated by a task coordinator to schedule up to 150 sub-agents for parallel retrieval and validation, consolidating the results into a shared evidence pool. A global validator audits all assertions by reasoning through the evidence graph to ensure that every conclusion in the final report is verifiable. Leveraging a multi-agent collaborative architecture, Apodex-1.0-H achieved a score of 90.3 in the BrowseComp web navigation test and a score of 60.8 in the text-only HLE complex reasoning test.
The training process is based on the Qwen3.5 series model, using supervised fine-tuning, direct agent preference optimization, and long-sequence reinforcement learning. While retaining the base model's general capabilities and code-level optimization, the development team has significantly increased the execution ceiling of deep research. Currently, the Smol series lightweight models with 0.8B, 2B, and 4B parameters have been open-sourced on Hugging Face, and the workflow runtime AgentOS and evaluation framework AgentHarness are also available on GitHub.
