DeepSeek Ships V4.1-Flash: 552B MoE Architecture Cuts HBM Memory to a Quarter
AI寒武纪 · wechat · 2026-09-10
DeepSeek released DeepSeek-V4.1-Flash, the smallest model in its new architecture family, with native vision understanding, faster inference, higher throughput, and scalability to larger models.
Architecturally, it has 552B total parameters in an MoE design using a new causal encoder-decoder architecture: only 8B active parameters when processing input and 16B when generating output. With a new pretraining method and larger-scale RL post-training, its benchmarks directly surpass the previous flagship V4-Pro. On KV cache, HBM memory drops to a quarter of the prior generation and SSD storage to an eighth — a major cost cut since cache-hit fees dominate agent usage costs.
V4.1-Flash is live on the DeepSeek API under the name deepseek-flash with native multimodal support. Since it beats the old flagship on performance, cost, speed, and total runtime, DeepSeek is phasing out V4-Pro: V4-Flash and V4-Flash-Vision-Exp are retired, with old endpoints temporarily routed to V4.1-Flash. From 04:00 UTC on September 14, 2026, all deepseek-v4-pro requests auto-switch to V4.1-Flash at the lower rate until V4.1-Pro launches. Peak/off-peak pricing remains, with off-peak at 50% of peak; new pricing takes effect 04:00 UTC September 10, 2026.
Weights and the tech report are open-sourced, and DeepSeek opened a channel for large-scale deployments with 2,000 GPUs and storage clusters.
More from Infra
- RTX 5090 hits $7K as PC gaming community goes into 'denial' over AI boom — PAstynome · 2026-10-27
- Wan2GP lands on Pinokio: 1-click AI video on 6GB GPUs — cocktailpeanut · 2026-10-27
- AWS on agent prompt caching: explicit + implicit patterns to slash token costs — AWS ML Blog · 2026-10-27
- Thieves spoofing GPS steal $100M+ in data center cargo since June — Ars Technica AI · 2026-10-27
- Lam Research: weekly token usage up ~250x in 18 months — Beth_Kindig · 2026-10-27
- AWS shows how to build enterprise agentic retrieval on Bedrock with CDK — AWS ML Blog · 2026-10-27