Baseten Engineer Decompiles Kimi K3 in 48 Hours: From 22,580 GPT-2s to KDA Evolution

A Baseten engineer spent 48 hours dissecting Kimi K3's code, revealing 22,580 GPT-2s inside, and tracing the evolution from GPT-2 to linear attention, DeltaNet, gated DeltaNet, to Kimi's own KDA, highlighting the unmatched transparency of open source.

• Kimi K3
• GPT-2
• DeltaNet
• 开源分析
• 模型架构演进