DeepSeek ships V4.1-Flash-1017 and opens Search and Fetch APIs for agents

DeepSeek released a new checkpoint of its V4.1-Flash model, named V4.1-Flash-1017, and made it live on its website, app and API. The company says it is faster and leaner than the September release: stronger scores, quicker task completion and shorter outputs, which it describes as an improvement in both performance and cost at once. It reports a score of 77.0 on the DeepSWE coding benchmark, and the weights appeared on Hugging Face in FP8.

Alongside the model, DeepSeek opened two standalone agent tools on its API platform to every developer. Search queries the web and returns structured results. Fetch retrieves a single page and returns it as Markdown. Both use the same authentication as the rest of the DeepSeek API and are priced at 5 yuan per 1,000 calls, so agent builders no longer need a separate search provider. DeepSeek also pushed a new release candidate of its open-source agent runtime, DeepSeek Harness, with a plugin index and more integrations.

The pattern is clear: DeepSeek is no longer just shipping cheap models. It is building the full agent stack around them.

Our take

A cheap coding model plus first-party web tools behind one key is exactly what cost-sensitive agent builders want. Western labs charge more for each of those pieces separately.