This edition covers 7:15 am on 17 October to 7:15 am on 18 October 2026, Sydney time. It records two changes: a DeepSeek model checkpoint update and the opening of standalone web-search and page-fetching APIs to developers.
DeepSeek releases V4.1-Flash-1017 and downloadable weights
On 17 October, DeepSeek updated V4.1 Flash on its website, app and API to V4.1-Flash-1017. The preceding version was formally released on 10 September. This is a new post-training checkpoint, not a repeat of the earlier API-routing change. The official update keeps the deepseek-flash model name, calling method and pricing unchanged.
The official Hugging Face repository lists weight files. Its model card places the repository and weights under the MIT licence and supplies prompt-encoding and local-inference references. No Jinja chat template is included; developers need the appropriate encoding implementation. Downloadable weights do not establish successful deployment on arbitrary hardware.
DeepSeek reports that DeepSWE v1.1 improved from 74.2 to 77.0 using the Mini-SWE-V2 harness. Reduced task duration and output length are also developer-reported. The Decoder corroborated the release channels and weight availability, while noting that independent benchmark results were not yet available. These figures are not presented as a universal capability ranking, and this review did not reproduce local inference. Direct references are the official changelog, model card and The Decoder report.
DeepSeek opens standalone Search and Fetch APIs
DeepSeek also opened its in-house search capabilities as standalone APIs that day. Search returns structured web results for a query; Fetch retrieves one page and returns Markdown content. Both use a DeepSeek API key, rather than requiring developers to obtain web information solely through a chat interface. This is a separate interface-access change alongside the checkpoint update.
Official documentation prices each API at CNY5 per 1,000 successful requests that produce results. Failed requests, timeouts and searches with no results are not charged. Search and Fetch each have an account-level limit of five requests per second. Fetch remains subject to the target website's availability and access policies; it does not guarantee access to every page.
This review checked public API documentation and independent release reporting, without making paid authenticated test calls. Search coverage, relevance and claimed quality advantages were not independently measured here. Returned scores are not treated as guarantees that content is correct. Direct references are the official API guide and The Decoder report.
Sources
DeepSeek; Change Log, entries for 17 October 2026 and the original 10 September release; updated 17 October 2026.
https://api-docs.deepseek.com/updates/
DeepSeek; DeepSeek-V4.1-Flash-1017 model card and weight repository; released 17 October 2026, checked 18 October 2026.
https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017
DeepSeek; Web Search and Fetch guide; accompanying the 17 October 2026 API release, checked 18 October 2026.
https://api-docs.deepseek.com/zh-cn/guides/search/
The Decoder; Deepseek V4.1 Flash 1017 beats larger rivals in agentic tasks while once again massively cutting per-task costs; 17 October 2026.
https://the-decoder.com/deepseek-v4-1-flash-1017-beats-larger-rivals-in-agentic-tasks-while-once-again-massively-cutting-per-task-costs/
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.