DeepSeek Ships a Faster Flash Model and Its Own Search API

DeepSeek is doing what it usually does: shipping more performance with very little ceremony. On Saturday, the Chinese A.I. lab rolled out DeepSeek-V4.1-Flash-1017, a new version of its fast Flash model, and said it finishes tasks faster, produces shorter answers and posts better benchmark numbers. The pricing did not move.

The update is already live in the web app, the mobile apps and the API under the name deepseek-flash. The biggest gains show up in coding and so-called agentic tasks, where the model has to work through several steps on its own. DeepSeek says Terminal-Bench 4.0 jumped from 31.2 in the September version to 42.3, while DeepSWE rose from 74.2 to 77.0.

Under the hood, V4.1 Flash is a mixture-of-experts model with 552 billion parameters, but only 8 billion to 16 billion are active per token. It has a one-million-token context window. Pricing stays at $0.30 per million input tokens and $1.20 per million output tokens, so shorter answers should make real-world use cheaper even if the sticker price stays put.

DeepSeek also opened its search engine to developers for the first time. The company says the system indexes hundreds of billions of web pages and handles hundreds of millions of queries a day, while trying to filter SEO spam, GEO manipulation and low-quality A.I. content. Through the new API, developers can fetch search results or ask for individual pages as text, and a thousand successful requests cost $0.75.

That puts DeepSeek in the same territory as Brave, Exa and Tavily, which already sell search APIs for A.I. agents. But the benchmarks need a grain of salt: the new version has only been measured by DeepSeek so far, and it is still unclear whether the company will publish the weights for the new checkpoint. The old September version was open on Hugging Face under an MIT license, but even DeepSeek users have learned one awkward truth — the model behind deepseek-flash can change without much warning.