DeepSeek opens its in-house search engine to developers, upgrades Flash model

Key Points

1. DeepSeek said it is opening API access to its in-house search engine powering web and API search.
2. The company also shipped DeepSeek-V4.1-Flash-1017, keeping the same pricing and API calling method.
3. Opening the index gives developers a search layer built for models rather than human browsing.

The latest:

DeepSeek is making its in-house search engine available through its API, the company said, describing a system that handles hundreds of millions of requests daily and indexes hundreds of billions of web pages. Alongside it, the company released DeepSeek-V4.1-Flash-1017, an upgrade now live on its web and app products as well. Pricing is unchanged.

Details:

• The search engine: DeepSeek described DeepSeek Search as an AI-native general-purpose engine that indexes hundreds of billions of web pages, supports multilingual queries and returns results in real time. The company said it is already the backbone of DeepSeek’s own web and API search features rather than a newly built product.
• Spam filtering: According to the company, the engine applies strict detection mechanisms to filter SEO spam, so-called GEO manipulation and low-quality AI-generated content. DeepSeek framed this as serving both human users and large language models, which are increasingly exposed to machine-written pages scraped at retrieval time.
• The model update: DeepSeek V4.1 Flash has been upgraded to DeepSeek-V4.1-Flash-1017, with the official web and app models moved to the new version. The API calling method and pricing remain unchanged, and developers access the latest build through the existing deepseek-flash identifier.
• Reasoning scores: The company reported 42.6 on HLE without tools, rising to 44.0 on the pure-text subset of the benchmark set. It also listed 35.4 on Agents’ Last Exam and 72.1 on Apex Bench v1.1, benchmarks aimed at long-horizon agentic tasks rather than single-turn answers.
• Coding benchmarks: DeepSeek published 42.3 on Terminal-Bench 4.0, 77.0 on DeepSWE v1.1, 68.2 on NL2Repo-Bench and 29.1 on ProgramBench. The company noted Terminal-Bench was run with the Claude Code framework and DeepSWE with Mini-SWE-V2, meaning results depend partly on third-party harnesses.
• Vision results: On multimodal tests run with tools, the company listed 92.5 on BabyVision, 83.0 on Chartography and 51.0 on ZeroBench-main, with the latter evaluated at Pass@5. Automation-Bench v1.0.6 came in at 60.5.
• Cost claim: DeepSeek said it cut task runtime and output length while raising scores, which it presented as a new performance–cost Pareto optimum across multiple harnesses. Shorter outputs translate directly into lower token spend for developers running the model at volume.
• Integration notes: The company said it optimized specifically for DeepSeek Harness, Tencent CodeBuddy, WorkBuddy and other platforms, and advised upgrading to DeepSeek Harness 0.2.1-rc.1 or later for best results with the new model. Usage instructions were pointed to the Search & Fetch guide.

Between the lines:

Two announcements landing together point to one strategy. A search index that filters AI-generated junk is most valuable to agents that browse autonomously, and the Flash upgrade is scored heavily on agentic benchmarks such as Agents’ Last Exam and Terminal-Bench. Selling retrieval and inference as one stack is a different business from selling model access alone.

What’s next

Watch whether independent evaluations reproduce the reported Terminal-Bench and DeepSWE scores outside the Claude Code and Mini-SWE-V2 harnesses, and whether DeepSeek publishes separate pricing for search API calls.