DeepSeek's official site now states that DeepSeek-V4.1-Flash has been updated to version V4.1-Flash-1017, with what it calls significantly enhanced project development capabilities, completing tasks at lower cost and faster speed. The same notice says the API now supports web search and fetch. The API reference lists two new endpoints: POST /v1/tools/search, which returns a ranked list of structured web results from an in-house search engine, and POST /v1/tools/fetch, both marked NEW. A changelog entry dated October 17 lists benchmark scores for the new build — HLE 42.6 without tools, Terminal-Bench 4.0 at 42.3, DeepSWE v1.1 at 77.0 — with footnotes naming the harness but no variance or per-task cost. The original V4.1-Flash announcement (September 10, 2026) described an asymmetric causal encoder-decoder architecture with 8B active parameters for input and 16B for output, a smaller KV cache, native visual understanding, and availability under the model name deepseek-flash.
Our takeA score table is not an eval protocol: the changelog names the harnesses but reports single numbers, so 'significantly enhanced' still can't be reproduced from the notes alone. The more consequential change is Search/Fetch — it turns the API from a completion endpoint into a retrieval loop the model can drive, which is where agent reliability is actually won or lost. Worth watching whether cache-hit billing and result ranking stay stable across builds.