DeepSeek released DeepSeek-V4.1-Flash on September 10, 2026, with architectural changes documented in the official model card. The backbone grew from 284B to 552B parameters while active parameters fell from 13B to 8B for input reading and 16B for output generation, reflecting a shift to a Causal Encoder-Decoder archit Benchmark jumps versus V4 Flash include Terminal-Bench 2.1 moving from 82.7 to 90.6, Terminal-Bench 4.0 from 7.0 to 31.2, and DeepSWE v1.1 from 54.4 to 74.2. DeepSeek announced that from 12:00 Beijing time on September 14, 2026 (04:00 UTC), requests to deepseek-v4-pro are routed to V4.1 Flash and billed at Flash rates,

DeepSeek V4.1 Flash is described as a low-cost, open-weight, multimodal model for long agent loops, combining a 1M-token context window with up to 384K output tokens and DeepSeek's official API ID deepseek-flash. The model's practical identity is five parts: multimodal image-and-text input, text output, a 552B-paramete Legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp temporarily route to V4.1 Flash at Flash rates, and from 04:00 UTC on September 14, 2026, DeepSeek routes requests for deepseek-v4-pro to V4.1 Flash until a later V4.1 Pro release. The piece explicitly recommends clients use the exact ID shown by their pro

DeepSeek V4.1 Flash moved from a two-day beta (carrying the internal ID deepseek-v4.1-flash-expires-on-0910) to a full general release on September 10, 2026, available across DeepSeek's app, web interface, and API under the model name deepseek-flash. DeepSeek published a technical report titled Pushing the Limits of KV The release artifacts include 48 safetensors shards totaling roughly 510 GB that went online the morning of release, letting developers inspect the new architecture directly. The piece also references the prior V4 Flash line and lists competing models served on the same routing platform, providing release-timing contex