Open-Source Implementation of Test-Time Compute Scaling and Agent Swarms

A developer has released an open-source project called Deepthink that implements test-time compute scaling and agent swarms using open-weights models like Gemma and Qwen. The system achieved massive performance improvements on difficult math and software engineering benchmarks. This project democratizes cutting-edge AI techniques by showing that smaller open-weights models can achieve dramatic performance gains through test-time compute scaling. It offers the community a transparent, accessible alternative to proprietary frontier reasoning models. The system processes about 200 million tokens per problem, with execution times ranging from 50 minutes to 30 hours depending on inference speed and depth. Benchmarks using Gemma and Qwen models showed score increases from 14% to 65% on Matharena and 9% to 52% on DeepSWE problems.