Chinese company DeepSeek, whose model V4.1-Flash-1017 holds the world's top spot for request volume as of November 2026, has decided to significantly tighten terms for free users of its application and web interface. According to official developer statements, the move is driven by a lack of computing power and record load growth: in the week from November 9 to 15, requests to the flagship model reached 59.2 trillion tokens. As a result, the company has implemented three key restrictions affecting millions of users who access the service daily without payment.

Three restrictions all free users have felt

Based on widespread online feedback and notices received by users, the new measures boil down to three specific actions. First, the inability to continue long dialogues that have shown no activity for more than seven days; some users received warnings 72 hours before automatic conversation closure. Second, a reduction in the maximum size of a single text request to approximately 40,000 characters, making it impossible to upload large documents in one go. Third, the introduction of a daily quota on the number of messages: once exhausted, the system displays a "server busy, usage quota exceeded" message. Additionally, users report that access can be restricted after just two hours of continuous use, though the company has not published exact quota algorithms or announced timelines for lifting these restrictions.

59.2 trillion tokens per week: How DeepSeek reached the top

Data from the OpenRouter API aggregator platform records explosive load growth. Version V4.1-Flash-1017, updated on October 17, took the top spot in the rankings as early as the week of October 19–25 with 33.6 trillion tokens — 801% more than the previous week. The earlier V4.1 Flash model took second place with 18 trillion tokens. Remarkably, all top five spots in the rankings during that period were occupied by Chinese large language models. By the week of November 9–15, requests for V4.1-Flash-1017 nearly doubled to 59.2 trillion tokens, allowing the model to maintain its global lead. For comparison: the model launched on September 10 reached 15.8 trillion tokens during its first full week (starting September 14). Thus, in just two months, the load increased nearly threefold, while infrastructure expansion apparently failed to keep pace.

Eight outages in a month: A chronicle of service degradation

DeepSeek’s official status page confirms that at least eight incidents have occurred since October, affecting the web version for end users, mobile app chats, and the API for V4 Pro and V4.1 Flash models. The most extensive was the October 15 outage, lasting over four hours, which limited access to both the user interface and the API for the two flagship models. This incident was the largest of the previous month. Numerous users between October and November reported "insufficient computing power" and "usage quota exceeded" notices, indicating systemic issues rather than isolated technical glitches.

Full API and open source: What remains for developers

Importantly, the implemented restrictions primarily affect the consumer version of the app and web interface. They do not apply to the API used by developers: it features separate limits on concurrent requests and operates on a pay-as-you-go model. On November 13, DeepSeek officially opened access to a full API, providing all app features—web search, analytical "deep" thinking, file and link handling, and voice input. Simultaneously, the company released the source code for the chat interface (chat-demo) under an MIT license, which, according to the developers, allows for creating a "full copy of the DeepSeek application." Theoretically, users can use the API to build their own service, though in this case, computing resources are no longer free.

Tiered rates and resource isolation: How DeepSeek manages load

To smooth out peak loads, the company has implemented several measures. Specifically, tiered rates have been introduced for the API depending on the time of day: costs are half as low during off-peak hours, encouraging developers to schedule background tasks for nighttime. Separate limits on concurrent connections have also been established—2,500 sessions for the deepseek-flash model and 500 for v4-pro, with these resources completely isolated from computing power allocated to regular app users. In essence, DeepSeek is reallocating capacity: the company is sacrificing some long-context capabilities to ensure service stability for active users. This primarily impacts tasks requiring ultra-long context, such as processing large documents, analyzing massive datasets, and multi-step analytical chains. The company did not announce when these restrictions might be reviewed or lifted as of the time of publication.