DeepSeek Releases V4.1 Flash 1017 Model on New CED Architecture With 1M Context

DeepSeek released the October 17, 2026 snapshot of DeepSeek V4.1 Flash (1017), a 552B-parameter sparse mixture-of-experts model and the first built on the company's Causal Encoder-Decoder (CED) architecture. The model uses an asymmetric split that activates 8B parameters on input and 16B on output to keep per-token compute low, and DeepSeek reports that it exceeds the 1.6T-parameter V4 Pro on performance, speed, and task completion time. Priced at $0.15 per 1M input tokens, $0.60 per 1M output tokens, and $0.003 per 1M tokens for cache reads, V4.1 Flash 1017 supports a 1,048,576-token context window and up to 393,216 completion tokens. Compressed KV caching reduces cache memory to roughly a quarter of the previous Flash generation, cutting costs for coding, terminal, and computer-use agents across long-horizon multi-step workflows. Image understanding is native to the CED architecture, with visual and text embeddings trained jointly from the start of pre-training rather than added afterward as in the earlier experimental V4 Flash Vision Exp.