IQuest Research releases IQuest-Q1 technical report
Start here
ZhiZhi Innovation Research Institute publicly released the complete technical report for IQuest-Q1, disclosing the 30 Benchmark results of this open-source base model with 320B total parameters and 15B activated parameters. The report states that the model scores 52.7 and 51.8 on AutomationBench and τ³-Banking respectively, ranking first among the comparison models listed in the report; training used approximately 29T Tokens and over 1 million containerized executable tasks, and suppresses Reward Hacking through a three-layer mechanism of environment cleanup, execution checks, and trajectory auditing.
Generated by AI from reports · Updated 20 days ago
Related research
Papers whose work the reports mention
- The reports mention CyberGymCyberGym: Evaluating AI Agents' Real-World Cybersecurity Capabilities at ScaleZhiZhi Innovation Research Institute publicly releases the complete IQuest-Q1 technical report: million-level task training and anti-cheating mechanisms for a 320B open-source base model
arXiv · 2025-06-03 · Deep reading
Progress timeline
- XiaohongshuXiaohongshu · Tech Just Now· DiscussionIQuest-Q1 public technical report(opens the original post in a new tab)
- WeChat · AI Era (Xinzhiyuan)ZhiZhi Innovation Research Institute publicly releases the complete IQuest-Q1 technical report: million-level task training and anti-cheating mechanisms for a 320B open-source base model
IQuest Research publicly released the complete technical report for IQuest-Q1, disclosing the 30 Benchmark results of this open-source base model with 320B total parameters and 15B activated parameters. The model scored 52.7 and 51.8 on AutomationBench and τ³-Banking respectively, ranking first among the comparison models listed in the report; training used approximately 29T Tokens and more than 1 million containerized executable tasks, and three layers of mechanisms—environment cleanup, execution checks, and trajectory auditing—were used to suppress Reward Hacking.
Related discussion
- Older than 72 hours · Not included in the 72-hour figures
- WebAI Era (Xinzhiyuan)1 million real-world problems, and AI cheating must also be prevented! Technical details of the 320B domestic base model revealed.(opens the original post in a new tab)
- XiaohongshuXiaohongshu · Tech Just NowIQuest-Q1 public technical report(opens the original post in a new tab)