Liang Wenfeng moves beyond the chat box: DeepSeek limits long conversations, opens API and open-sources frontend
Since November 15, multiple DeepSeek users have reported that the mobile app shows a notice that long conversations exceeding 72 hours are temporarily unavailable due to insufficient computing power, while another user found in testing that a single session triggers the limit when it accumulates to 148,000 characters; the company has not yet commented. Previously, on November 13, DeepSeek opened capabilities including web search, deep-thinking search, file input, link reading and voice as APIs, and open-sourced the chat-demo chat frontend under the MIT license, supporting the import of historical conversations. On this basis, the article argues that Liang Wenfeng regards the chat box as a basic form, with the focus placed on AGI research and the DSH plugin ecosystem.
Why read this
AI score 63
The article ties together DeepSeek's limiting of long conversations, opening of its API and open-sourcing of its frontend as one thread, and explains its trade-offs by comparison with the history of Apple's platforms.
Topics
Body · AI translation
On the afternoon of November 15, multiple DeepSeek users suddenly reported that when they tried to continue a conversation from several days earlier, the mobile app displayed a prompt: "Due to insufficient computing power, long conversations exceeding 72 hours are temporarily unavailable. You can export the conversation to use in other apps."
In the community, some users compiled several other error messages that appeared at the same time, such as "Content too long, please shorten and try again," "Input modification count exceeds limit, message not sent," and "Message count has reached the limit, please start a new conversation."
One user tested it and said that a single conversation of theirs triggered the limit when it accumulated to 148,000 characters.
At present, the feedback is mainly concentrated on the mobile app. As of press time, no official statement from DeepSeek has been seen.
In fact, there were "omens" of this.
On November 13, a developer dug out a complete set of quota management logic in the front-end code of chat.deepseek.com, with the prompt text: "Usage quota has reached the limit, please try again after your quota is restored."
From then on, the news kept spreading within the community, and by the 14th, it had been circulated that DeepSeek was going to charge membership fees.
On the evening before the news spread, DeepSeek released the "full suite application API," opening up all the capabilities in the app—web search, deep-thinking search, file input, link reading, and voice—as interfaces, and also open-sourced a chat front end called chat-demo.
DeepSeek officially also said that as long as you use this set of tools, you can build a DeepSeek app of the same kind yourself.
On one hand, handing out tools so developers work for it; on the other hand, setting limits on its own product—isn't this Jobs-like?
More like dispersal than harvesting
According to the community's compilation, when the DeepSeek web page is visible, it syncs the quota of each model every 5 minutes; when the quota is exhausted, a prompt bar appears above the input box. Based on this, the community speculated that DeepSeek was going to launch a plan called Coding Plan, together with V4.1 Pro.
But these speculations are actually a bit excessive. Some users checked its code item by item and found that there was actually no plan pricing, no subscription entry, no payment channel, or anything like that.
There is indeed an "upgrade" in the code, but it only reminds you that a new version of the web page is available; you click here and the page refreshes and upgrades. It has nothing to do with paying to upgrade.
Looking back at the prompt on the 15th, when other companies' free quotas run out, what usually pops up is a button to become a member. What DeepSeek pops up is an urging for you to export the conversation and then take it to another app to use.
A company, within its own product, points users to a way out. As expected, it has to be Liang Wenfeng.
In its announcement on the 13th, DeepSeek wrote that developers can use this set of APIs to "replicate a complete DeepSeek App 1:1."
At the same time, the open-sourced chat-demo uses the MIT license, requires no backend, and can be used by filling in your own API Key. The announcement specifically stated that it "supports directly importing DeepSeek historical conversations."
After the same update, DSH also gained a conversation import plugin, so historical conversations from the app and web version can be imported into the workspace and continued.
chat-demo created its repository on the 12th, open-sourced the front end, opened the interfaces, and connected imports on the 13th, and long conversations became unavailable on the 15th. Although DeepSeek has not explained the direct connection between these events, it first gave you an exit and only then set the limit.
There is also a sentence at the end of the announcement: "DeepSeek's existing conversation product is only a relatively basic form."
This echoes what Liang Wenfeng said before. He once said: "The most important reality last year was that everyone had to make Chatbots and compete for consumer traffic. But we do not think it is an important matter, or rather, within our company, what we truly care about is actually the AGI roadmap I just mentioned, and how this technology will break through next."
DeepSeek currently has about 130 million monthly active users. But even so, in the official narrative it is merely a "basic form."
On November 8, Cui Tianyi posted on social media. He used smartphones as an example: the system has built-in alarm clocks and calculators to meet users' basic needs, but manufacturers that make operating systems hope even more that better alarm clock apps and calculator apps than the built-in ones will appear in the ecosystem.
He then wrote in the post that he hoped DSH could become the operating system of the Agent era.
According to Cui Tianyi's explanation, the chatbot is the built-in alarm clock and calculator in this system. DeepSeek only needs to do it about well enough, which is why it was open-sourced and handed over to the community.
There is a saying in Apple's developer circles called "Sherlocked."
This saying originated in 2002, when Apple added new features to its search tool Sherlock that highly overlapped with third-party software Watson, causing the means by which Watson survived to be taken away by built-in system features.
Liang Wenfeng did the opposite. DeepSeek split its built-in app into interfaces, and developers took them and handed them to third parties.
Computing power is the number one problem
That said, the reason given in the prompt is "DeepSeek insufficient computing power."
So someone questioned on the forum: "Who would talk about the specific problems of their own backend on their own product?"
But this itself was Liang Wenfeng's judgment. Liang Wenfeng once said: "What stops us from scaling is actually computing power, not that we don't want to scale."
At an earlier investor exchange meeting, he attributed the gap between Chinese and American AI to one thing: "All the differences we see, including differences in talent, differences in model capabilities, and differences in applications, can be considered to all stem from differences in computing power resources." As for domestic computing power, his judgment was that there is no problem with either hardware or the ecosystem, "the only problem is that production capacity is insufficient."
There are only so many chips, and production capacity cannot be made up anytime soon, so isn't the only remaining question "who gets to use them first"?
So Liang Wenfeng's logic is: "The first goal of the models we make is not for everyone to use them well, but for us to use them well ourselves. After they are useful to us, I will be faster when developing the next version of the model."
Liang Wenfeng has stated two standards. One is the main line: DeepSeek only does AGI, and "both C-end and B-end are byproducts on our path to AGI." The other is not subsidizing: when DeepSeek V2 cut prices in 2024, he said his principle was "no subsidies, and no excessive profits."
At the very front is certainly research itself, because it is the main line of AGI itself. At this point, if you look back at Liang Wenfeng open-sourcing chat and opening up these APIs, it makes sense again.
After all, DeepSeek has no "physical body"; it is just a layer of software, and for computing power it can only rely on others.
But computing power has not stopped DSH's advance. According to data disclosed by DeepSeek officially, as of November 4, the daily active users of DSH using DeepSeek's official API had reached 1.54 million, consuming about 92 trillion tokens per day.
There are about 35,000 third-party plugins, and about 60% of active users have made or installed at least one plugin themselves. Yet DSH's v0.1 only went online on August 13, and the desktop version was only released on September 29. In its first week, daily active users were only 271,000.
Only making the operating system and a few basic capabilities, then bringing developers from all over the world to write software for it, and using a store to collect the results—this is the typical Apple-style playbook.
On March 6, 2008, Apple released the iPhone SDK. On July 10, the App Store opened, with 500 apps on its shelves; developers got 70% of revenue and Apple got 30%. Jobs himself was not responsible for developing apps; he played the role of platform legislator.
DeepSeek has also announced that it plans to build an official plugin marketplace.
Liang Wenfeng is not really "driving" users away. What he wants to be is the Jobs of the Agent era.
At WWDC in 1997, Jobs had just returned to Apple not long before. On stage he said that focus is not saying "yes" to the thing you should focus on, but saying "no" to another hundred good ideas. "I'm as proud of the things we haven't done as the things we have done."
Liang Wenfeng once told a 36Kr reporter that DeepSeek does not do verticals and applications, but does research and exploration.
Only by first making your own territory smaller is there room for others to stand. Putting a limit on the chat box is DeepSeek saying "no."
In the middle of DSH is a plugin system called Cordis, responsible for managing plugin loading, unloading, and dependencies. The 53 built-in tools that DSH ships with are all plugins, and even the Web interface is a plugin—meaning you can even replace the official UI.
DSH's later-released creation mode lets people create a plugin on the spot using only plain language. This gives people who cannot write code, for the first time, the qualification to participate in defining the system. Ultimately, all of this is about lowering the cost of developing DSH plugins, and thereby making more people willing to develop them.
On March 6, 2008, Apple released the iPhone SDK. Just 4 days after release, it was downloaded more than 100,000 times. 95 days later, on June 9, 2008, the official figure was that downloads had exceeded 250,000.
Just two months after the App Store opened, there were more than 3,000 apps. Later, by April 2009, there were 35,000 apps on the App Store, with cumulative downloads of 1 billion.
The same 35,000: the App Store took 9 months, while DSH took 83 days.
Of course, one plugin does not carry the weight of one app, and besides, making plugins on DSH can also be done by an Agent. But Liang Wenfeng is indeed retracing the path Jobs walked, and walking it faster.
Will DeepSeek trigger an App Store moment?
The essence of the App Store moment is not the store, but the rules. The whole industry is forced to redo software according to your rules. What the iPhone changed was not only users and developers, but technology itself.
Still on that stage on January 9, 2007, Jobs mocked the stylus: "Who wants a stylus? You have to get them out, put them away, and you lose them. Yuck. Nobody wants a stylus."
Immediately after, he said: "We're going to use the best pointing device in the world, the one we're born with, the one we're born with ten of—our fingers."
Jobs's logic was that if users are made to adapt to technology, they have to read manuals and spend time learning how to use it. Apple wanted technology to adapt to people's bodies, people's language, and people's lives. You don't need to learn; you know how to use it from birth.
DSH's creation mode lets people build plugins in plain language, and the thinking is similar. However, although DSH already has 1.54 million daily active users, it is still a little short of triggering an "App Store moment."
The most direct gap is monetization.
DeepSeek's support program currently offers only three things: token funding, technical exchange, and promotion of works. There is no pricing, no revenue sharing, and no payment channel. At this stage, the official side has only released a form to collect expressions of interest, explicitly stating that participation rules and support details will be announced only after the official launch.
On this point, Apple did much better back then.
Steve Demeter originally wrote ATM software at a large bank. In 2008, he used his evenings and weekends to develop a game called Trism, which caught the App Store launch and was one of those 500 apps, priced at $4.99.
It was a match-three puzzle game, with triangular tiles that could be slid in six directions; tilt the phone, and the tiles would also fall with gravity. It was different from the gem-matching games common at the time and was designed specifically for the iPhone's hardware features.
In the first two months on the store, Demeter earned $250,000 from Trism. He quit his bank job and started his own game company. Once the news spread, more developers rushed into Apple's operating system to seek their fortune and build all kinds of apps.
People need a sense of honor to live, but people cannot live on honor alone; they also need to eat.
In addition, developers also need a "DeepSeek development guide."
DeepSeek officially said only that it would improve the stability and compatibility of the plugin API and strengthen plugin security, without publishing any review process. The official side itself also admitted in a long post that community plugins are still evolving and vary in maturity and applicable scenarios.
This order is actually reversed; review should come first, then installation. The "armor-breaking" plugin dsh-purge in the DeepSeek community is one example.
DSH itself carries a layer of host-side protection, including approval interception and system prompts. In many cases, when the AI refuses to answer, it is not the model itself refusing; the Harness layer blocks it first.
The logic of dsh-purge is to replace and clean the Harness's built-in safety system prompts in the local session, separating host interception from the model's native output, so that the model can output prohibited content in DSH that would otherwise be blocked. The author positioned it as a red-team evaluation tool, while the community simply calls it "armor-breaking."
In late October, the original author announced that it would be closed source, and the original repository was once deleted, but mirrors and forks had long since spread out, and the version taken over by the community is still being updated to this day. If DSH had reviewed before publishing, such plugins could have been kept out and would not have affected the community's reputation.
But review means intervention, which conflicts with what Liang Wenfeng said about "as little intervention and management as possible."
However, Liang Wenfeng does not necessarily have to follow Jobs. The "DSH moment" is far more impactful than "yet another App Store moment."
This article comes from the WeChat public account "Letter AI," author: Miao Zheng, published by 36Kr with authorization.