bubble boi
The argument that deepseek cuts KV down more is moot, because every layer has KV and it grows wit...
20 hours agobubble boibubbleboi
Twitter
View on X
Note: AI-generated summary based on third-party content. Not financial advice. Read more.
Quick Insights

The technical architecture of Kimi K3 and DeepSeek V4-Pro is driving a shift toward tiered data assets, increasing the importance of DRAM as a warm-cache layer and NAND for large-scale historical session storage. Despite DeepSeek reducing its KV cache to 10% of previous versions, channel checks indicate that NAND usage has actually increased in real-world deployments due to higher offload ratios. Moonshot recommends deploying the 2.8T parameter Kimi K3 on high-bandwidth supernodes with at least 64 accelerators, maintaining high demand for HBM, GPUs, and scale-up networking.

Ask about this postAnswers are grounded in this post's content.
Tweet
About bubble boi
bubble boi

bubble boi

By bubbleboi

head of risk @thru_xyz.