DeepSeek Releases V4.1-Flash-1017 With Live Web Search

What happened

DeepSeek, the Chinese AI company, released an updated version of its V4.1 Flash model on 17 October 2026. According to the company's own API change log, the new build is named DeepSeek-V4.1-Flash-1017 and replaces the earlier V4.1 Flash behind the existing deepseek-flash name. The company describes it as a minor update that improves reasoning, coding and agent tasks. The weights were posted to Hugging Face under an MIT licence, with a copy on ModelScope, the model hosting platform run by Alibaba's ModelScope community.

The web search piece arrived the same day, but it is a separate service. According to CoinDesk reporting carried by KuCoin, DeepSeek launched Search and Fetch APIs that supply real-time web content to AI agents and retrieval applications. Coverage gives the price as $0.75 per 1,000 successful requests, with a cap of 5 requests per second per account. One outlet reports the fee as 5 yuan per 1,000 requests, so the figure should be checked against DeepSeek's current price page before budgeting.

Why it matters for businesses that automate work

Most automated office tasks break down when the software needs information that is not in its own files. A supplier's stock page, a customs rule or a shipping tracker sits outside the business. A model that can fetch live pages can complete that step without a person copying data across. DeepSeek is not the first company to offer this, but it is one of the lower-cost options reported this month, and the open weights give technical teams a second route: running the model on their own hardware.

The benchmark claims come from DeepSeek. According to MarkTechPost, the new build beats the previous V4.1 Flash on all 11 benchmarks that DeepSeek reported, by margins of 2.0 to 11.1 points. Those results have not been independently tested in the coverage reviewed here, so treat them as the company's own measurements.

What changes for a business that automates work

The practical effects depend on how the business already uses DeepSeek or a similar model. The following points apply to firms of any size:

• If you call the API under the deepseek-flash name, you receive the new version without changing code. The company says the calling method and pricing structure stay the same.
• Search is a new cost line. According to aiweekly, the Search and Fetch endpoints bill per successful request, and paid billing starts on 31 October 2026 at 16:00 UTC. Coverage from TechInsider also reports the same start date.
• Shorter outputs and shorter run times lower the cost per task, according to the company's model card. Measure this on your own tasks, not on the benchmark tables.
• Self-hosting is a hardware decision. MarkTechPost reports that the weights are split into large files and that hardware requirements were not disclosed for this build. Hugging Face lists the repository at 763 billion parameters, which needs servers with a lot of GPU memory.

The parameter count needs care. The 552 billion figure is the backbone size DeepSeek gives. MarkTechPost adds a 196 billion parameter memory module, and Hugging Face's repository label is 763 billion. The two numbers measure different things, so they should not be compared directly.

One limit is worth stating. A model that reads the live web can still act on a wrong or outdated page. Any automated step that moves money, changes a supplier record or contacts a customer should keep a person in the approval loop.

What to watch next

Three dates and signals matter over the next few weeks. First, paid search billing starts on 31 October 2026. According to aiweekly, searches triggered through Claude Code and Codex are free until 1 November 2026, so teams using those tools should check their invoices after that date. Second, the model briefly appeared in DeepSeek's API list before the official release, according to Startup Fortune, so the official change log is the reliable reference. Third, look for independent tests of the 1017 build. Until those exist, the open question is how the model performs on the specific tasks a business runs every day.