Community
Developer open-sources Deepthink, a test-time scaling system tested on Gemma4-35B and Qwen-3.8-27B
The author reports scores rising from 14% to 65% on 15 BrokenArxiv problems and from 9% to 52% on 8 DeepSWE problems, using about 200M tokens per problem. Runs take 50 minutes to 30 hours depending on depth and inference speed, and the claims have not been tested on frontier models.