MaxKernel system was released on September 7, 2026. Its performance was significantly better than that of the baseline according to JaxBench evaluation.
2026-09-07 12:00Models🔥 47.2 heat score
2sources
1days unfolding
47.2heat score
7mentions
SummaryAI generated
On September 7, 2026, the research team released the MaxKernel system. This multi-agent framework is designed to automate TPU kernel development. The system offers three operating paradigms: human-machine collaboration (HITL), fully automated optimization, and graph-based global search, and shares sub-agents for planning, implementation, debugging, testing, and hardware analysis. After evaluation with 50 diverse tasks and real open-source model workloads using JaxBench, MaxKernel generated optimization results comparable to those of expert manual tuning baselines, and achieved significant performance improvements in benchmark tests. The project has been opened sourced, and the code is hosted on the GitHub repository.
MaxKernel system was released on September 7, 2026. Its performance was significantly better than that of the baseline according to JaxBench evaluation.
Coverage · reports per dayLANGUAGE SPLIT
Entity relations
Integrated timelineUNIFIED TIMELINE
2026-09-07
Release and evaluation of MaxKernel system
The research team released the MaxKernel multi-agent framework, offering three paradigms: human-machine collaboration, full automation, and graph search. Evaluation conducted on 50 tasks and real-world workloads showed results comparable to those of the expert baseline, with significantly improved performance.
MaxKernel is a multi-agent system that automates TPU kernel development through collaborative, autonomous, and graph-searching paradigms. The system includes three modes: a human-machine collaborative design agent, a fully automated metric-driven optimization agent, and an extensible global exploration graph-searching agent. These three modes share specialized sub-agents for planning, implementation, debugging, testing, and hardware analysis. The research team evaluated MaxKernel on complex real-world workloads using JaxBench (with 50 diverse kernel tasks) and state-of-the-art open-source models. The results showed that the optimized implementations generated by MaxKernel were consistent with those achieved by expert manual tuning, and they delivered significant performance improvements in benchmark tests. The project is now open source, and the code can be found at https://github.com/AI-Hypercomputer/accelerator-agents/tree/main/MaxKernel.