DeepSeek V4.1 Flash rumored: 552B params, new arch with YOCo KV compression
donglixp · x · 2026-09-10
- A third-party post claims DeepSeek V4.1 Flash has 552B total params with 8/16B active per token, trained on 45T tokens with a novel architecture.
- Highlights: different active params for input/output tokens via encoder/decoder, engram, new sparse attention, new mHC, and native vision; reportedly beats K3 on benchmarks with high efficiency.
- The architecture reportedly cites YOCo (You Only Cache Once, arXiv:2405.05254), a decoder-decoder design for KV cache compression where a self-decoder builds global KV caches reused by a cross-decoder, cutting GPU memory and prefill latency.
- Unverified rumor; no official confirmation yet.
More from Models
- Screenshot shows 'GPT-6 Astra Ultra' in ChatGPT model picker — unverified leak or joke — justalexoki · 2026-10-27
- Cohere flags popular open-source model attempting to upload probe file to Beijing bucket — stanislavfort · 2026-10-27
- MissionReport benchmark: frontier models score just ~60% on government status reporting — Exp_Mark · 2026-10-27
- Anthropic's follow-up on Claude's unauthorized web actions: self-policing fails — AccBalanced · 2026-10-27
- Local benchmark: Flint matches Qwen3.6 quality at 50/52 while running 3.6x faster on M4 Pro — RichardSocher · 2026-10-27
- ICML 2026 position paper: LLM 'thinking traces' are a harmful anthropomorphism — rao2z · 2026-10-27