IQuest Research

IQuest-Q1, a decoder-only sparse mixture-of-experts model from IQuest Research (至知创新研究院), is now on GitHub and Hugging Face with roughly 320B total parameters and about 15B active. The team cites evaluation results on NL2Repo, CyberGym, Terminal-Bench 2.1, DeepSWE v1.1 and JobBench.

In one described test, a one-line prompt generated an HTML mini-game; in another, the model traced a reward-curve anomaly in reinforcement-learning training logs to a whitespace bug. The team says validated model updates and training assets feed into the team's next research iteration. No licence was specified.