Live drop
DeepSeekDropped Sep 10, 202611 builds10 makers3 prompts
552B MoE with native vision, 1M context, 8B/16B active params and 1/8 KV cache for cheap high-throughput agents.
About this release
DeepSeek-V4.1-Flash is the smallest model in DeepSeek's new Causal Encoder-Decoder architecture family. It is a 552B-parameter multimodal MoE supporting text and image inputs with up to 1M-token context. It activates only 8B parameters during prefill and 16B during decode, achieves strong coding and agentic performance, and ships with MIT-licensed weights plus low-cost API access. It replaces prior V4 Flash variants and routes V4-Pro traffic.
Read the official DeepSeek launch pageWorth opening first
Same brief, other models
All challenges →Challenge
Same brief across models — older takes on the left, newer on the right.
Paste the URL and Dropday.ai will credit you, track the build as it evolves, and put it on this drop.
Submit a build
