ModelsSep 17, 2026 · 18:521 source

DeepSeek V4.1 Flash Cuts Prefill Compute 50% in Shift From Standard Transformer

From huggingnews.com

Image: huggingnews.com
AI;DR

DeepSeek-V4.1-Flash uses a Causal Encoder-Decoder design with a 40-layer backbone split into a 20-layer causal encoder for prompt processing and a 20-layer decoder for token generation, reducing prefill compute overhead by roughly 50% compared to standard transformers. The architecture employs FP4 global KV to lower the cache to 890 bytes per token, uses a memory sharing system called CSA2 to reduce the KV cache, and incorporates mHC kernel optimizations to lower communication overhead between SM and HBM. Increasing reasoning effort from 25 to 100 raised the average Pass@1 across eight benchmarks from 67.1% to 76.3% and increased output length 2.5 times.

Read the sourceSign in to suggest

Sources

#deepseek#llm#inference#transformer#reasoning#benchmark#v41#flashMarkdown
Content history (2)
  • 2026-10-15 · Summary · vi · Replaced stored text
    DeepSeek-V4.1-Flash sử dụng thiết kế Causal Encoder-Decoder kích hoạt 8 tỷ tham số cho mỗ…
    DeepSeek-V4.1-Flash sử dụng kiến trúc Causal Encoder-Decoder, kích hoạt 8B tham số cho mỗ…
  • 2026-10-15 · Title · vi · Replaced stored text
    DeepSeek V4.1 Flash giảm 50% tính toán prefill trong chuyển đổi từ Transformer tiêu chuẩn
    DeepSeek V4.1 Flash cắt giảm 50% tính toán prefill trong chuyển đổi từ Transformer tiêu c…

One ranked edition, every morning

The stories that matter today. Free, cancel anytime.

Telegram @aidr_today