DeepSeek has released DeepSeek-V4.1-Flash-1017, a new checkpoint of its V4.1-Flash model that is currently live on the company's website, app, and API. The updated model delivers stronger benchmark scores, faster task completion, and shorter outputs compared to the September release, achieving 77.0 on DeepSWE v1.1 and outperforming models like Claude Opus 5 in specific evaluations. Alongside the model update, DeepSeek opened its Search and Fetch APIs to all developers, providing a first-party web layer and a cheaper coding model accessed through a single API key.
The V4.1-Flash architecture utilizes a multimodal Mixture-of-Experts design featuring 40 Transformer layers split into a 20-layer causal encoder and a 20-layer decoder. Innovations such as Compressed Sparse Attention 2, an FP4 main KV cache, and SWA Bounded Replay help reduce computational overhead while maintaining a 1-million-token context window. The newly opened Search and Fetch APIs allow developers to query ranked web results or retrieve single pages as Markdown, backed by an in-house engine priced at $0.75 per 1,000 successful requests with server-side calls remaining free until November 1, 2026.
This release matters because it lowers the cost and increases the capability for agent builders by integrating a high-performing coding checkpoint with native web search tools under competitive pricing tiers.