DeepSeek V4.1 Flash
Günstig mit optionalem Denkmodus
Best for: math, logic, code and low-cost everyday questions · fast
- No login required
- Understands images
- Web search
- Thinking mode
approx. 227 tokens per response
DeepSeek · 1 model
DeepSeek
DeepSeek is inexpensive to use and suited to math, logic and code. Start right here in your browser, without an account with the provider and without login.
Genuine DeepSeek models through the official API, no account with the provider. As a guest you get 15,000 tokens with DeepSeek V4.1 Flash. A free account adds 15,000 tokens a day, all models, uploads and saved chats.
What is DeepSeek?
DeepSeek is an AI lab from Hangzhou, China, that grew out of the hedge fund High-Flyer in 2023 and releases its language models with open weights. It made its name in January 2025 with the reasoning model R1, which kept up with the big US models in math and code while running far more cheaply. On ChatX you use the current model, DeepSeek V4.1 Flash, through the official API, with no DeepSeek account, no app and no phone number.
The DeepSeek models on ChatX
Choose the model that fits your task. Fast models are enough for everyday questions, while the larger models pay off for long texts, analyses, and code.
New at DeepSeek
Blog →
DeepSeek Can Now Search the Web
DeepSeek V4.1 Flash now has web search on ChatX. We asked the same question without and with web search and show how many tokens it uses.
DeepSeek V4.1 Flash reads images
DeepSeek has released V4.1 Flash, a small model that understands images and beats the larger V4 Pro in tests, so clearly that DeepSeek is r…
Privacy: what happens to my input?
Your input goes from ChatX straight to DeepSeek's official API. The operator is Hangzhou DeepSeek Artificial Intelligence Co., Ltd., based in China, and the provider's privacy policy names China as the location of its servers. What DeepSeek receives through the API is the text of your request along with the previous chat history. According to DeepSeek's privacy policy (as of February 2026), data is stored in China and, once personal identifiers are removed, may be used to a limited extent to improve the services; there is no commitment not to use API data for training. The proceedings by European data protection authorities against DeepSeek (Italy 2025, Berlin 2025) concern the app and the chatbot service, not the API.
You don't need a DeepSeek account, an app or a phone number. On ChatX your chats are kept only in your own account, and you can delete them at any time. Deleting them on ChatX, however, doesn't change what the provider stores under its own terms. So don't enter confidential or personal data you don't want to share.
On politically sensitive topics related to China, the model answers in a limited or evasive way; this is built into the model itself and doesn't affect everyday questions, math or code.
DeepSeek compared with the other models on ChatX
Pick DeepSeek when you want to do math, code or solve logic problems cheaply and care about thinking mode: it's strong here and uses far fewer tokens than ChatGPT, Claude or Grok. For creative writing with a fine feel for language, Claude is ahead, and for up-to-the-minute questions you switch on web search, which DeepSeek runs without thinking mode; on political topics involving China, DeepSeek's answers are limited. On ChatX you switch between all five providers in the same chat and compare the answers side by side.
| Model | Strength | No sign-up | Features |
|---|---|---|---|
| ChatGPT | All-rounder for writing, analysis and code | GPT-6 Luna | Web search · Understands images · Thinking mode |
| Claude | Long texts, careful language, code | Claude Haiku 5.5 | Web search · Understands images · Thinking mode |
| Gemini | Fast, large context, understands images | Gemini 3.5 Flash Lite | Web search · Understands images · Thinking mode |
| Grok | Direct tone, quick answers | Grok 4.3 | Understands images · Thinking mode |
| DeepSeek | Math, logic and code at a low price | DeepSeek V4.1 Flash | Web search · Understands images · Thinking mode |
All models run in the same chat. Switch any time without losing your history.
Open source: run DeepSeek yourself?
DeepSeek publishes the weights of its models under the MIT license, on Hugging Face among other places. That means anyone can download them, run them on their own hardware and build them into products. For companies with strict privacy requirements, this is the way to use DeepSeek without requests leaving the building; European hosting providers offer the models as well.
For private users, that's rarely practical. The large models need computing power no laptop has; the small, stripped-down versions do run locally but give noticeably weaker answers. That's why the easy route for most people is to use DeepSeek through an API, just like here on ChatX, with no DeepSeek account and no app.
If you still want to try it, you'll find the models under the name DeepSeek on Hugging Face, and the license also allows commercial use. What the model does or doesn't say about political topics related to China won't change, though; that is built into the model itself.
What you can do with DeepSeek on ChatX
Math and calculations
DeepSeek works through problems step by step, and in thinking mode it handles tricky word problems too.
Code
Write code, find bugs and get someone else's code explained, from Python to JavaScript.
Logic and puzzles
Logic problems, brainteasers and reasoning with a solution you can follow.
Write texts
Emails, summaries and short texts in your tone, at low token usage.
Translate
Carry texts over between English, Chinese, Spanish and other languages.
Learn and understand
Get school subjects and specialist topics explained clearly, with examples and quiz questions.
DeepSeek, explained simply
DeepSeek is an AI model from China that is among the low-cost models on ChatX; you can compare usage per reply in the model overview. It is suited to math, logic and programming. The current model DeepSeek V4.1 Flash comes with an optional thinking mode, in which it reasons step by step before replying, and with an account it understands images as well as text. Using it directly with the provider requires an account; on ChatX you use DeepSeek through the official API without an account with the provider and without signing up, with an English interface and billing per reply. DeepSeek understands English and replies in the language you write in. Your requests are sent to DeepSeek's API to be answered; what the provider stores is governed by its terms, so do not enter confidential data.
It's this easy
-
1
Enter your question
Type your question or task into the input field. You don't need to log in for this.
-
2
Choose a model
Start with the default model or switch to a more powerful one anytime.
-
3
Keep going
With a free account, you save your chats, upload files, and get free tokens every day.
Costs and free allowance
- Without signing up, you get 15,000 tokens to try it out. That's enough for around 66 answers with DeepSeek V4.1 Flash.
- With a free account, you get 15,000 tokens as a gift every day.
- If you need more, you can buy tokens with no subscription or choose a monthly plan. Billing is based on the model's actual usage.
Frequently asked questions
Is it legal to use DeepSeek?
Yes, using it as a private individual is not banned. In 2025, EU data protection authorities criticized the DeepSeek apps over data transfers to China: Italy's regulator stopped the processing of Italian users' data, and Berlin's data protection commissioner reported the apps to Apple and Google; bans such as Australia's apply only to government devices. On ChatX you use DeepSeek without the app and without an account at the provider, but the advice about sensitive data still applies.
Is DeepSeek censored?
On politically sensitive topics related to China, the model dodges or answers briefly. For math, code, translations and everyday questions, this makes no difference.
Does DeepSeek have a thinking mode?
Yes. You switch thinking mode between Off and High next to the model selection, even without signing up. DeepSeek V4.1 Flash then works through the problem step by step before answering, which helps with math, logic and code; the thinking phase counts as output and costs correspondingly more tokens.
Is DeepSeek open source?
DeepSeek publishes its model weights openly, for R1 under the MIT license, and they have been downloaded millions of times. Running it yourself only makes sense with a lot of hardware; on ChatX you use the current model without a computer of your own.
Does DeepSeek speak English?
Yes. You write in English and get answers in English, no language setting needed, and the same goes for many other languages. The model handles dialects and slang less well than standard English.
Can DeepSeek understand images?
Yes, with a free ChatX account: DeepSeek V4.1 Flash reads uploaded photos, screenshots and document pages and answers questions about them. Without signing up, uploads are not enabled.
What is DeepSeek less suited to?
You can switch on web search for up-to-date questions with DeepSeek, but it then answers without thinking mode. For very creative writing it comes across as plainer than Claude or ChatGPT.
What does DeepSeek cost on ChatX?
DeepSeek is among the low-cost models on ChatX. You can see the approximate usage per answer in the model overview; without signing up you start with a starting allowance, and with a free account you get free tokens every day.
The history of DeepSeek
From the research group of a Chinese hedge fund in 2023, through the global R1 shock in January 2025, to an agent and search platform with funding in the billions in 2026.
Newest events first
-
V4 and agents · 2026
-
Firma
Agent framework Harness reaches 1.5 million daily users
DeepSeek reports 1.54 million daily active users of its open agent framework Harness for early November and around 92 trillion tokens processed per day. The figures come from the company, but they show that a software ecosystem of its own is forming around the models.
1.54M users/dayAbout 35,000 pluginsCompany figures
-
Firma
About 100 billion yuan in new capital reported
Reuters reports that DeepSeek is close to closing a funding round of about 100 billion yuan at a valuation of around 500 billion yuan, with investors including Tencent, CATL and Meituan. As of the cutoff date, DeepSeek itself had not confirmed the details.
About 100 bn RMBValued at about 500 bn RMBNot confirmed
-
Produkt
Own web search opened up, V4.1 Flash updated
DeepSeek makes its in-house search engine available to developers as the Search & Fetch API and speaks of hundreds of millions of searches per day. On the same day, an updated V4.1 Flash arrives with a focus on coding and agent workflows. Since November 2026, this web search can be switched on for DeepSeek V4.1 Flash on ChatX.
Search & Fetch APIV4.1-Flash-1017
-
Modell
DeepSeek V4.1 Flash is released
The new Flash model understands images as well as text, answers faster and remains the latest main model generation as of the cutoff date. On ChatX it is the guest model, DeepSeek V4.1 Flash, with an optional thinking mode.
552 bn parameters (MoE)MultimodalOn ChatX
-
Firma
Preparing for a possible IPO
According to Reuters, DeepSeek hires CITIC Securities to prepare a listing on the STAR Market in Shanghai. As of the cutoff date there is no prospectus, approval or date.
STAR Market ShanghaiNo date
-
Modell
V4 Pro for everyone, Harness as an open agent framework
The large V4 Pro becomes generally available on the web, in the app and through the API, along with agent features and control over thinking effort. At the same time, Harness launches, an open framework in which every capability is a plugin.
V4 Pro generally availableAdjustable thinking effortHarness preview
-
Firma
First outside funding round
After years in which the hedge fund High-Flyer alone backed the lab, DeepSeek brings in outside investors for the first time: more than 50 billion yuan, from Tencent and CATL among others. Later investor documents point to a valuation of around 351 billion yuan.
> 50 bn RMBTencent, CATL
-
Regulierung
Italy's competition authority closes case with commitments
The AGCM accepts commitments from DeepSeek, including clearer warnings that the model may make false statements. The case ends without a finding of a violation.
AGCMHallucination warning
-
Modell
DeepSeek V4 preview with a one-million-token context
The fourth generation arrives as V4 Pro and V4 Flash with a context window of up to one million tokens and multimodal capabilities. According to Reuters, DeepSeek worked closely with Huawei on it, and parts of training and operation run on Ascend chips.
1M token contextV4 Pro 1.6 trillion parametersHuawei Ascend
-
Firma
OpenAI accuses DeepSeek of model distillation
In a memo to US lawmakers, OpenAI accuses DeepSeek employees of getting around access restrictions and using output from its models for training. DeepSeek has not commented, and no court has examined the allegation.
Allegation, no ruling
-
R1 and the world stage · 2025
-
Modell
DeepSeek V3.2 and the reasoning variant Speciale
At the end of the year, V3.2 arrives together with V3.2 Speciale, a variant focused on long reasoning. This finally ends the era of separate chat and reasoning models.
V3.2V3.2 Speciale
-
Forschung
Over 86 million downloads of the open models
A report by the US institute NIST counts more than 86 million downloads of DeepSeek models on Hugging Face, up from about 4 million a year earlier. The open weights have become a standard building block worldwide.
86M downloadsHugging Face
-
Modell
V3.1 combines chat and thinking mode
With V3.1, DeepSeek introduces hybrid reasoning: one model either answers right away or thinks step by step first. It also brings better capabilities for tools and agents, and the Terminus update follows in September.
Hybrid reasoningTool use
-
Regulierung
Berlin reports the DeepSeek apps to Apple and Google
Berlin's data protection commissioner classifies the transfer of user data to China as unlawful and reports the apps to the app store operators under the Digital Services Act. Before that, she had asked DeepSeek to make changes or withdraw the apps.
DSA noticeData transfer to China
-
Modell
R1 gets a major update
The R1-0528 version clearly improves the reasoning model on logic and programming tasks and keeps pace with the competition.
R1-0528
-
Regulierung
South Korea demands privacy fixes
The data protection authority PIPC finds that DeepSeek transferred user input abroad without sufficient consent, including to the ByteDance subsidiary Volcano Engine. DeepSeek stops these transfers and is given conditions; before that, new app downloads in the country had been suspended.
PIPCVolcano EngineDownloads paused Feb 2025
-
Forschung
Open Source Week: the tools behind the models
For a week, DeepSeek releases a piece of infrastructure every day, such as FlashMLA, DeepEP and the 3FS file system. The lab shows how it keeps training and operation so cheap.
FlashMLADeepEP3FS
-
Regulierung
Australia bans it on government devices, German regulators investigate
Australia bans DeepSeek on government systems, and in Germany several state data protection authorities launch a coordinated review over the lack of an EU representative and data transfers. There is no ban for private users in Germany.
Government devices onlyCoordinated review in Germany
-
Regulierung
Exposed database and Italy's privacy stop
Security researchers at Wiz find an unprotected database with over a million entries, including chat histories; DeepSeek quickly closes it, and no misuse has been proven. At the same time, Italy's data protection authority bans the processing of Italian users' data.
Wiz discoveryGarante ItalyCyberattack Jan 27
-
Firma
No. 1 in the US App Store and the Nvidia shock
The DeepSeek app overtakes ChatGPT as the most downloaded free app in the US. On the same day, Nvidia loses about 17 percent, or almost 600 billion dollars, on the stock market, as investors stop believing in the AI industry's hunger for computing power.
No. 1 US App StoreNvidia down 17%
-
Modell
DeepSeek R1: the breakthrough
The reasoning model R1 is released with open weights under the MIT license and, according to DeepSeek, reaches the level of OpenAI o1. On the same day, founder Liang Wenfeng attends a meeting with Chinese Premier Li Qiang; the consumer app had launched a few days earlier.
MIT licenseOn par with OpenAI o1App since Jan 10
-
V2 to V3 · 2024
-
Modell
DeepSeek V3: big, efficient, open
V3 has 671 billion parameters, of which only 37 billion are active per word. The paper cites about 2.8 million GPU hours for training, a fraction of what was assumed for comparable US models.
671 bn parameters37 bn active2.8M GPU hours
-
Modell
DeepSeek V2 lays the foundation
With V2, DeepSeek relies for the first time at large scale on the DeepSeekMoE expert architecture and the memory-saving Multi-head Latent Attention. In September, the chat and coder models merge into V2.5.
DeepSeekMoEMLAV2.5 in Sept.
-
Founding · 2023
-
Modell
The first models: DeepSeek Coder and DeepSeek LLM
Within a month, the young lab releases its first family of code models and its general language models with 7 and 67 billion parameters, both with open weights. The focus on programming still shapes DeepSeek today.
Coder 1B to 33BLLM 7B and 67BOpen weights
-
Firma
Founded in Hangzhou
Hangzhou DeepSeek Artificial Intelligence Co., Ltd. is registered, with Liang Wenfeng as founder and, to this day, CEO. Back in March, his hedge fund High-Flyer had announced it would build its own research group for general AI.
Liang WenfengHigh-FlyerHangzhou
Compiled from public announcements by DeepSeek and press reports; updated when new models arrive.
More AI models on ChatX
Last reviewed: 10. November 2026 · Source: DeepSeek