Does anyone know the proportion of post-training/RL steps of o1-mini, o3-mini and o4-mini compared to each other? In the o3 and o4-mini release statement (https://openai.com/index/introducing-o3-and-o4-mini/) they say that ‘we’ve pushed an additional order of magnitude in both training compute and inference-time reasoning’. Does anyone have more detailed information about the scaling of the post-training time of these models?
Related topics
| Topic | Replies | Views | Activity | |
|---|---|---|---|---|
| Announcement: Release of o3 and o4 mini - April 16, 2025 | 15 | 6085 | April 17, 2025 | |
| When will we get reasoning models + search + RAG + code execution? | 0 | 310 | January 26, 2025 | |
| OpenAI dropped price of o3 | 21 | 2527 | June 17, 2025 | |
| GPT-4o mini slow inference | 6 | 793 | April 9, 2025 | |
| O1-PREVIEW reasoner is using neuro-symbolic AI? | 3 | 407 | December 10, 2024 |