DeepSeek has released and open-sourced the V4.1 Flash-1017 model, achieving significant breakthroughs in agentic and coding tasks. Priced at just $0.30/$1.20 per million tokens, it scores 45 on the Artificial Analysis Intelligence Index, ranking second among open-weight models. Its DeepSWE benchmark score of 77 approaches the closed-source flagship Opus 5, yet costs only 1/83 as much.
Technically, V4.1 Flash employs an asymmetric Mixture-of-Experts (MoE) architecture with a Causal Encoder-Decoder (CED), activating only 8B/16B parameters out of 552B total. Combined with KV cache compression that reduces memory usage by ~75%, it achieves Pareto improvements in performance and cost. The introduction of Self-Feedback Reinforcement Learning (RLSF) enhances end-to-end project development capabilities without human evaluation. Cross-verified data shows it tops the Code Arena WebDev leaderboard while shortening output lengths.
This release reshapes the high-value AI ecosystem. With MIT licensing and availability across all channels, it lowers barriers for developers. Native support for multimodal and long-horizon agentic tasks makes it ideal for complex applications. As platforms like Databricks and OpenRouter integrate it, V4.1 Flash accelerates enterprise adoption, challenging the notion that open-source cannot compete with closed-source leaders.