This retweet summarizes DeepSeek V4's key technical breakthrough in inference efficiency. The primary contribution is a 10x reduction in KV cache requirements at 1M context length compared to V3.2 (requiring only 10% of…
DeepSeek released DeepSeek-V4 Preview as open-source, featuring two variants: DeepSeek-V4-Pro with 1.6T total parameters and 49B active parameters, and DeepSeek-V4-Flash with 284B total parameters and 13B active paramete…
DeepSeek disclosed a RoPE (Rotary Position Embedding) implementation mismatch bug in the DeepSeek-V3.2-Exp inference indexer module. The core issue: Indexer RoPE expects non-interleaved input tensors while MLA RoPE expec…
