# OpenAI's major results over the last 3 months came from scaling test-time compute with agent swarms. I've been building an open-source equivalent for the past year, and I'd love for you guys to try it. — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/c708bd1d48a81264

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-10-27 20:47 ingested 2026-10-28 02:38

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1xh4bnd/openais_major_results_over_the_last_3_months_came/)

## retained text



```
I have been working on test-time scaling for quite some time now. Posted some of my previous work on this sub-reddit [before](https://www.reddit.com/r/LocalLLaMA/comments/1u47cvc/i_scaled_testtime_compute_for_qwen3627b_and/). Recently, I made significant simplifications to the core system-design and it finally felt like now anybody can understand and directly use this on their hardest problems. Just on a side-note though, depending on the depth you set & your LLM's inference speed, it can take from 50 Minutes to 30 hours to give you back the results.

  
For benchmarks I couldn't test this on frontier models or even large open models because it will be way too expensive. I directly only tested this on Gemma4-35B & Qwen-3.8-27B. They spent around 200M tokens / problem and there scores went from 14% to 65% on BrokenArxiv (Matharena) 15 problems or from 9% to 52% on 8 DeepSWE problems. I know that's almost unbelievable, but you can try and see that yourself.

  
Repo Link: [https://github.com/ryoiki-tokuiten/Deepthink](https://github.com/ryoiki-tokuiten/Deepthink)
```



document fetched 2026-10-28 00:42 · canonical https://reddit.com/r/LocalLLaMA/comments/1xh4bnd/openais_major_results_over_the_last_3_months_came
