Commenter speculates the model was distilled from Opus and benchmark-maxed later
LucaAmb · x · 2026-07-23
A commenter suggests the model may have been distilled from Opus and then further benchmark-maxed through post-training.
They add that post-training fine-tuning can be extremely time-efficient, implying that much of the apparent jump in performance could come after the base distillation step.
More from Models
- Merge says every Fusion setup beats solo frontier models at one-quarter the cost — shensi · 2026-07-23
- Merge launches Fusion, a one-API multi-model judge that claims 4× lower cost — shensi · 2026-07-23
- Kimi says Chinese is its native tongue, but English still handles most tasks well — amyxlu · 2026-07-23
- Chinese AI models now account for 60% of token usage at U.S. companies — SumitGup · 2026-07-23
- A user asks whether Kimi reasons better in Chinese than in English — amyxlu · 2026-07-23
- Developer Finds Claude 3 Opus Agents Stuck in Subagent Forking Loop — bdsqlsz · 2026-07-23