DeepSeek Ships V4.1-Flash: 552B MoE With 8B Active, 75% Less KV Cache Memory
mark_k · x · 2026-09-11
DeepSeek released DeepSeek-V4.1-Flash, the smallest model in a new architecture family, claiming it beats V4-Pro on performance, cost and speed. It's live on the API and open-sourced on Hugging Face, with V4-Pro requests switching over September 14 ahead of a future V4.1-Pro.
Key specs and efficiency gains:
- 552B-parameter MoE with only 8B active for input processing and 16B for output
- Native visual understanding
- KV cache uses just ¼ the GPU memory and ⅛ the SSD storage of the previous generation — a big deal for agents on long tasks
markk notes: if this is the smallest model in the family, the upcoming V4.1-Pro should be significant.
More from Infra
- EdgeRock open-sourced: 72 MiB KV cache at 256K context, 99.1% smaller than Qwen3.5-4B — alifcoder · 2026-10-27
- EdgeRock-4B holds KV cache at 72 MiB across 256K context, 99.1% smaller than Qwen3.5-4B — alifcoder · 2026-10-27
- RTX 5080 price jumps from $1658 to $2318 in ten days on Amazon — Frizzy-MacDrizzle · 2026-10-27
- ComfyUI on 12GB VRAM: measured peaks show offloading turns 24s into 11 minutes — Impossible-Mail-4619 · 2026-10-27
- Neon rebuilds its docs search entirely on its own stack: Lakebase Search, Functions, AI Gateway — matei_zaharia · 2026-10-27
- PyTorch details 2.12x faster on-device streaming ASR on Arm SME2 via ExecuTorch — PyTorch · 2026-10-27