Eigen RadarAI
Archive
Analysis

DeepSeek opens its search tools alongside the new Flash model

DeepSeek has released V4.1-Flash-1017 and made its search and webpage-fetching services available to outside developers. The update targets software projects that require sustained work across many steps. Its weights are distributed under the MIT license, while claims of better coding performance come from the developer’s own evaluations. The model is accessible through the web, app and developer interface.

Artificial Intelligence··Evening
Wooden desk, keyboard and grey computer enclosure in a bright room.

The Flash update reaches developers

DeepSeek, the developer of the Flash model family, released V4.1-Flash-1017 on 17 October and opened its own search and webpage-fetching tools to outside developers. The model is available through the company’s website, mobile application and application programming interface, or API. Search and Fetch provide separate ways for software to request search results and retrieve a page’s contents. The services accompany the model revision rather than an announcement of a future demonstration.[1], [2]

The update centres on software work that continues across a whole project. DeepSeek says it used reinforcement learning with feedback the model generates about its own work to improve consistency on these tasks. That is the developer’s account of the training approach and its benefits. Its comparisons do not amount to an independent test establishing that the model performs better across all coding assignments.[1]

Weights and execution tools have separate roles

DeepSeek recommends pairing the model with DeepSeek Harness v0.2.1-rc.1, its framework for running agent tasks. The announcement also identifies adaptation work with Tencent’s CodeBuddy and WorkBuddy coding tools. Model weights are distributed on Hugging Face under the MIT license. The accompanying model documentation separates code that converts conversation formats from the software that actually performs inference: an interface conversion does not execute the model.[1]

The company’s own coding comparisons include DeepSWE v1.1 with the mini-swe-v2 execution framework. They describe task success alongside the cost of completing a task. Those measurements belong to DeepSeek’s evaluation, with the model release and the newly accessible tools forming the concrete changes announced to developers.[1]

Search calls and model access follow different arrangements

DeepSeek says server-side tool use through DeepSeek Harness, Claude Code and Codex is free through the end of October. The temporary offer concerns tool calls; general model API pricing remains a separate arrangement. Developers can therefore use the search and fetching functions alongside software-building workflows without treating the promotion as a change to every part of the model service.[1]

References

  1. News sourceDeepSeekDeepSeek releases V4.1-Flash-1017 and opens search and page-fetching APIs↩1↩2↩3↩4↩5
  2. News sourceDotAI TIMESDeepSeek’s Flash update opens search tools to outside developers↩
AI safety and regulationThe frontier model race

More analysis from this desk

View all Brief analyses from the Artificial Intelligence desk →