September 10, 2026: DeepSeek-V4.1-Flash
The smallest model in DeepSeek’s new architecture family, with native visual understanding: a 552B-parameter MoE using a Causal Encoder-Decoder design with just 8B active parameters for input and 16B for output, tuned for cheaper, faster inference.
What this means for your agents: Agent loops burn tokens; a cheaper multimodal model with real throughput cuts the cost of running agents daily.
June 24, 2026: DeepSeek-V4: million-token context
DeepSeek-V4 introduced highly efficient million-token context intelligence.
What this means for your agents: A million tokens is a full year of inbox exports or ledger history in one prompt - whole-archive agent passes become practical.
More DeepSeek agent resources
AI agent skills for DeepSeek - the honest setup guide for running our 28 agents on DeepSeek. Start with the Sourcing Agent on DeepSeek guide if you want a concrete example.
AI Agent Skills is an independent store. We are not affiliated with, endorsed by, or sponsored by DeepSeek or its makers. Release details belong to their sources; check DeepSeek's own documentation for the authoritative record.