Open-source Deepthink scales test-time compute, lifting 27B models from 14% to 65%
Ryoiki-Tokuiten · reddit · 2026-10-28
A developer released Deepthink, an open-source take on the test-time scaling + agent swarm approach behind OpenAI's recent results, with a much simplified system design.
- Tested only on Gemma4-35B and Qwen-3.8-27B (frontier models too expensive)
- Costs 200M tokens per problem; runs 50 minutes to 30 hours depending on depth
- Scores jumped from 14% to 65% on BrokenArxiv (Matharena) 15 problems and 9% to 52% on 8 DeepSWE problems
- The architecture diagram shows parallel mainline agents doing correction/critique iterations, global CONTINUE/PRUNE/EVOLVE strategy updates, plus a shared workspace board, self-improving verification suite, and git-style strategy history rollback
The author admits the numbers are 'almost unbelievable' and invites the community to reproduce them. Repo: github.com/ryoiki-tokuiten/Deepthink
More from coding & agent
- ai-design-skills rockets to 9.1k GitHub stars with 1,395 stars in a single day — elayadesign · 2026-10-28
- EmbeddingGemma2 Turned Into a 20ms Real-Time Trainable Multimodal Endpoint on CPU — michellechen · 2026-10-28
- Anthropic hiring manager kills the coding interview: "It no longer provides signal" — felixrieseberg · 2026-10-28
- Nous' memelord Goes Beyond Research, Now Hookable via MCP — alexcovo_eth · 2026-10-28
- Custom node pack brings looping H3 video generation to Comfy Cloud — PurzBeats · 2026-10-28
- Chaosfolk: open-source tool sics 16 AI personas on your web app to break it — Vjeux · 2026-10-28