Open-source Deepthink claims agent-swarm test-time scaling lifts small-model scores from 14% to 65%
Ryoiki-Tokuiten · reddit · 2026-10-28
A developer open-sources Deepthink, a year-long project replicating frontier labs' test-time compute scaling via agent swarms.
- Tested only on Gemma4-35B and Qwen-3.8-27B (frontier runs too expensive), burning 200M tokens per problem.
- Scores jumped from 14% to 65% on BrokenArxiv (Matharena) 15 problems, and 9% to 52% on 8 DeepSWE problems.
- The architecture includes Window Memory Folding (windowed critique-correct loops, distilled memory banks, global updates) and synchronous multi-branch control (parallel slots, BSR window reports and a standing ledger, pressure gate θ=0.20 driving continue/rollback/prune decisions).
- The author admits the numbers are "almost unbelievable" but says it's reproducible; repo is public.
More from coding & agent
- Adi Polak jokes about a 'CLAUDE_CODE_EVEN_MORE_DANGEROUSLY_SKIP_PERMISSIONS' env var — AdiPolak · 2026-10-28
- DeepSeek-V4.1-Flash Q2 runs at 41 tok/s on one RTX 3090, 90% token match vs API — musicurgy · 2026-10-28
- Resend launches Inboxes, an inbox for agents, as developer marks 3-month sprint — philnash · 2026-10-28
- Design team: AI now produces the work overnight, humans just make the calls — thesephist · 2026-10-28
- Chasing the Karpathy Peak: researcher maps AI productivity's slop cliff — anshulkundaje · 2026-10-28
- USearch v3 RC ships after 1,000 days, its libraries now embedded in 550K-star projects — viglovikov · 2026-10-28