DeepSeek used its October 17 API changelog to announce DeepSeek-V4.1-Flash-1017, an update to V4.1 Flash, the smallest model in its current architecture family. The company says its official web and app models now run the new version, and that the API calling method and pricing are unchanged - the existing deepseek-flash model name routes to the new checkpoint.
DeepSeek reports benchmark improvements over the previous V4.1 release, including HLE (without tools) at 42.6, Terminal-Bench 4.0 at 42.3, DeepSWE v1.1 at 77.0, NL2Repo-Bench at 68.2 and Automation-Bench v1.0.6 at 60.5. These are vendor-reported figures on the company's own changelog, with test harnesses noted for some entries: Terminal-Bench 4.0 was run with the Claude Code framework and DeepSWE with Mini-SWE-V2, and the HLE score used only the benchmark's pure-text subset. Independent replication was not available at publication.
The company also says it reduced task runtime and output length for the new checkpoint, and optimized it for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy; it recommends DeepSeek Harness 0.2.1-rc.1 or later.
The same changelog opens DeepSeek's in-house search engine to developers through new Search and Fetch APIs. DeepSeek describes it as an AI-native general-purpose engine indexing hundreds of billions of web pages with multilingual support and filtering for SEO spam and low-quality AI-generated content - the company's characterization, not an independent assessment.
A DeepSeek-V4.1-Flash-1017 model page is live on Hugging Face under the deepseek-ai organization. V4.1 Flash is the open-weight entry in DeepSeek's new architecture family, which the company says is designed for a higher capability ceiling, faster inference and scaling to larger models.
DeepSeek is a Chinese AI lab; the model and APIs are offered globally through its own API and web/app products. Regional availability of specific features was not detailed in the changelog.