If you still think of DeepSeek only as a “cheap but risky Chinese AI,” you may be missing a new option. DeepSeek attracted attention for its reasoning performance and the way it released its models. Now it is targeting tasks such as game development and 3D modeling.

Should you use it now? Try the AIB comparison card below to compare DeepSeek’s answers with those from the AI you normally use. Before entering customer information or unpublished work materials, check your company’s rules and the service’s data-handling terms. Even with the same DeepSeek model, your data takes a different path depending on whether you use the official app, another company’s service, or your own installation.

Here is why DeepSeek was so disruptive, what has changed in the latest V4.1 release, and what the privacy controversy actually involved.

Why DeepSeek Was Disruptive: Performance, Efficiency, and Open Weights

DeepSeek drew attention as an affordable AI. The key was delivering high performance at a lower training cost and releasing models that people could run themselves.

Released in late 2024, V3 combined an architecture that activates only the parts of the model needed for a question, rather than the entire model every time, with memory and training optimizations. It demonstrated a more efficient way to build and operate large AI models.

The V3 report estimated the cost of its official training run at $5.576 million. For a sense of scale, Stanford’s 2024 AI Index estimated GPT-4’s training compute cost at about $78 million. In a market where training costs in the tens of millions of dollars were being discussed, DeepSeek’s figure in the single-digit millions drew attention to efficiency. However, the two estimates use different calculation methods and training conditions. The V3 figure covers the official training run, excluding prior research and experiments; it is neither the company’s total development spending nor the cost of developing R1.

Released in January 2025, R1 focused on reasoning: working through problems step by step. In some math, coding, and reasoning evaluations published by the company, it reported results close to those of OpenAI’s o1 at the time. DeepSeek also released the model weights.

“Open weights” means releasing the “model weights,” the result of an AI model’s training. Put simply, you can download the trained AI and run it on a sufficiently capable computer or company server. Instead of having to access a hosted service, you can choose where and how to operate it. This does not mean that all the training data and development processes are public, and you still need to account for the hardware and costs of running it.

This does not mean R1 was better at every task. What matters is that individuals gained another option to compare, while businesses gained more choices over costs and how to run AI. DeepSeek raised a clear question: “Which AI capabilities and services are we paying for?”

The Latest V4.1-Flash-1017: Look at the Numbers at the End

As of October 19, 2026, the latest public update is DeepSeek-V4.1-Flash-1017. It went live on the official website, app, and API on October 17. The API model name is deepseek-flash. It is a follow-up improvement to the V4.1 family first released in September.

Read the final 1017 as a calendar date: October 17, the day this update was released. Think of it as a date stamp saying “October 17 edition.” If you look only at the first part of the name and assume it is the same model, you may miss what has changed.

The company says it used feedback from the model’s own evaluation of its results in reinforcement learning to improve its ability to complete projects from start to finish, including game development and 3D modeling. It also reported shorter task completion times and shorter outputs than the previous version.

Released alongside the update, Search finds web search results, while Fetch reads a specific webpage and retrieves its text. These tools let AI find up-to-date information as it works. See DeepSeek Search and Fetch API news for details on the terms.

The latest model’s code and weights were released under the MIT License. In plain language, you can download, use, and modify them, including for commercial services. When copying or distributing them, you must retain the original copyright notice and license notice.

Is DeepSeek Banned? What the Privacy Controversy Actually Involved

South Korea’s privacy investigation offers a concrete example for understanding the controversy. DeepSeek temporarily suspended new downloads from South Korean app stores on February 15, 2025, then resumed them on April 28. This was not a blanket block on existing apps and the website.

There were confirmed problems. South Korea’s Personal Information Protection Commission (PIPC) announced that users’ prompts had been sent to Chinese cloud provider Volcano Engine, and that notices and consent procedures for overseas data transfers were inadequate. The central issue was that users had not been sufficiently informed about where the text they entered was going.

However, claims that keyboard input patterns or rhythms were actually collected differ from what the regulator confirmed. Although this appeared in the privacy policy, the PIPC confirmed that the data had not actually been collected. It also confirmed that new transfers of prompts to Volcano Engine had been blocked since April 10.

Improvements still need to be checked. In a September 2026 follow-up review of privacy policy evaluations, the PIPC recommended that DeepSeek improve its processes so that requests to exercise privacy rights could be received and handled smoothly. The resumption of downloads does not mean that every data-handling issue has been resolved.

App, API, or Self-Hosting: The Same Model, Different Choices

Open-weight models can be offered by multiple providers or operated independently. This means the company that created the AI and the company running the service may be different. Even with the same DeepSeek model, the path your data takes and the applicable terms depend on whose servers run it.

Check the model developer, service operator, server and data storage locations, whether data is sent elsewhere, and the policies on retention and use for training. Having servers in your country does not, by itself, guarantee that all data stays there.

• Official website and app: The simplest starting point. Check the data-handling terms for the service operated by DeepSeek.
• Official API: A way to connect the model to your own software. Check the contracts and data-handling terms that apply to the API.
• Another company’s service: Do not rely solely on the statement that it uses a DeepSeek model. Check the actual operator and the path your data takes.
• Self-hosting: Run the publicly available weights in your own environment. You are responsible for managing external connections, logs, and server security.

The official service’s current Korean-language privacy policy explains how to opt out of use for training by turning off the “Improve the model for everyone” setting. This does not eliminate all data transfers and retention needed to provide the service.

Start with a simple comparison. Give DeepSeek and your usual AI the same request, such as turning a public product description into a customer response or fixing test code. Assess accuracy, wording in the language you will actually use, the number of revisions needed, and total cost together. This will reveal the differences that matter for your work.

Choose Based on Today’s Results, Not Yesterday’s DeepSeek

If you leave DeepSeek off your comparison list solely because of past controversies, you may miss a new option. If you switch based only on price, you may spend more time reviewing and correcting its work. Test whether the latest model’s strengths match the tasks you repeat.