DeepSeek Speeds Up V4.1-Flash With Minor Update, Launches Search APIs for Developers

Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co. Ltd., a Chinese artificial intelligence startup, has unveiled a refreshed iteration of its V4.1-Flash model designed to execute tasks with improved speed. The new DeepSeek-V4.1-Flash-1017 arrives just over five weeks following the company's initial V4.1-Flash launch, which debuted as the lightest offering within its new architecture lineup. According to its listing on Hugging Face, the revision represents a "minor update" that "reduces task execution time and output length," though DeepSeek has not disclosed specific performance gains.

When the original model debuted in September, affordability served as its primary draw, with off-peak API rates beginning at 15 cents per million uncached input tokens. Bloomberg reported that following this launch, equity values for Chinese competitors MiniMax Group Inc. and Alibaba Group Holding Ltd. declined in Hong Kong trading.

Testing results demonstrate incremental progress on DeepSeek's own benchmarks. The refreshed version resolved 77% of tasks on the DeepSWE v1.1 software engineering evaluation, compared to 74.2% for the September release. Anthropic PBC's Claude Opus 5 achieves 74% on the same assessment. However, Anthropic's model still leads across seven of the 11 published benchmarks, and maintains an advantage of nearly 14 points on Humanity's Last Exam.

DeepSeek optimized the update specifically for agent-based software, targeting its own DeepSeek Harness tool alongside two products from Tencent Holdings Ltd.: CodeBuddy, a coding assistant, and WorkBuddy, a desktop application for office tasks. WorkBuddy has supported V4.1-Flash since the model's September 10 release.

Beyond the model refinement, DeepSeek introduced additional capabilities for developers building agent applications. The company made its internal search engine accessible through a new Search API capable of returning up to 50 ranked web results per query. A complementary Fetch API transforms individual web pages into Markdown format. Both services carry a charge of 75 cents per 1,000 successful requests, with complimentary access through DeepSeek Harness, Anthropic's Claude Code, or OpenAI Group PBC's Codex until November 1.

These announcements follow a nearly five-hour service disruption on October 15 that Yicai Global identified as among DeepSeek's most severe outages this year. The company issued an apology and compensated users with credits reflecting their spending during the preceding week. DeepSeek committed to eliminating single points of failure within its infrastructure to prevent similar incidents. Bloomberg quoted certain analysts attributing the outage to insufficient computing resources.

The revised model is currently accessible through DeepSeek's web and mobile applications. API users can access it using the existing deepseek-flash model identifier at the same pricing structure, with the release notes recommending Harness version 0.2.1-rc.1 or later. Model weights are available on Hugging Face under the MIT license.