What Is DeepSeek AI?
DeepSeek AI is a Chinese AI lab, and the family of free and low-cost AI models it builds. Here is who is behind it, what it offers and how it got here.
DeepSeek is a Chinese artificial-intelligence lab based in Hangzhou that builds large language models and gives the public three ways to use them: a free chatbot (web and mobile), a low-cost developer API, and open model weights anyone can download and run.
It was founded in 2023 by Liang Wenfeng as an offshoot of the quantitative hedge fund High-Flyer. DeepSeek became globally known in January 2025, when its R1 reasoning model matched leading US models at a reported fraction of the training cost and its app briefly topped app-store charts worldwide.
Its signature is efficiency: mixture-of-experts models that switch on only a small slice of their parameters per token, sparse attention for long inputs, and aggressive API pricing. The current lineup is V4.1 Flash (552B parameters, native vision) and V4-Pro (1.6T parameters).
- Dec 2024DeepSeek-V3 makes low-cost MoE mainstream
- Jan 2025R1 reasoning model and app go viral
- Dec 2025V3.2 introduces DeepSeek Sparse Attention
- Apr 2026V4 family previews with 1M-token context
- Sep 2026V4.1 Flash adds native vision
- Oct 2026~$12B funding round reported ahead of a planned IPO
Want the long version? Read our full explainer: What is DeepSeek AI? →
Why Choose DeepSeek AI? 8 Key Benefits
Why so many people and developers pick DeepSeek over other AI chatbots and APIs (lower cost, strong coding, open weights and more), plus when it is not the right fit.
Much lower cost
V4.1 Flash costs a fraction of most frontier APIs, cached input is just $0.003 per million tokens, and off-peak hours are half price.
Frontier-level coding & agents
DeepSeek reports V4.1 Flash ahead of its own V4-Pro on agentic coding, with CyberGym 88.1 and DeepSWE v1.1 74.2.
Open weights you can own
Download and self-host the same models behind the API, fine-tune them on your data, and keep everything on your own servers.
Huge context window
Fit whole codebases, long contracts or book-length drafts in a single prompt, with a compact KV cache keeping long runs cheap.
Sees images natively
Send screenshots, charts, invoices or photos of homework in the same request as text, with no separate vision model.
Easy to switch to
Works with OpenAI- and Anthropic-format SDKs: change the base URL and model name, and your existing code keeps running.
Free to try
The official chat and mobile apps are free, so you can test DeepSeek on your own tasks before spending anything on the API.
Control speed vs. depth
Choose low, high or max reasoning effort per request to trade cost and latency against answer quality.
Output price per 1M tokens
Lower is cheaper · off-peak list prices
DeepSeek prices from its official pricing page (off-peak; peak is 2×). Claude Opus 4.8 list price as listed in our comparison table. Input pricing and real costs depend on your workload.
Maybe not the best pick if…
- You handle regulated or confidential data and can't self-host. The hosted service is run by a Chinese company.
- Your organisation restricts Chinese AI providers, or you work in US government or defence supply chains.
- You need answers on politically sensitive China-related topics; the hosted models filter them.
- You need audio input or image generation, which the current models don't offer.
All DeepSeek AI Models
A complete list of DeepSeek models, from DeepSeek LLM to V4.1 Flash, with release dates, sizes, context length and whether each is on the API or available as open weights.
| Model | Released | Type | Size (active) | Context | Status | Note |
|---|---|---|---|---|---|---|
| DeepSeek-V4.1-Pro | TBA | General · flagship | TBA | TBA | Upcoming | Confirmed in development, no date |
| DeepSeek-V4.1-Flash | Sep 2026 | General · vision | 552B (8B in / 16B out) | 1M | On API now | API: deepseek-flash |
| DeepSeek-V4-Pro (0813) | Aug 2026 | General · agents | 1.6T (49B) | 1M | On API now | API: deepseek-v4-pro |
| DeepSeek-V4-Flash-Vision-Exp | Aug 2026 | Vision · experimental | — | 1M | Open weights | API retired 10 Sep 2026 |
| DeepSeek-V4-Flash (0731) | Jul 2026 | General · fast | 284B (13B) | 1M | Open weights | API retired 10 Sep 2026 |
| DeepSeek-V4 (preview) | Apr 2026 | General | Pro 1.6T · Flash 284B | 1M | Open weights | Preview, replaced by GA builds |
| DeepSeek-V3.2 | Dec 2025 | General · reasoning | 671B (37B) | 128K | Open weights | Introduced Sparse Attention |
| DeepSeek-V3.2-Exp | Sep 2025 | General · experimental | 671B (37B) | 128K | Open weights | First DSA release |
| DeepSeek-V3.1 / Terminus | Aug–Sep 2025 | Hybrid thinking | 671B (37B) | 128K | Open weights | One model, thinking + non-thinking |
| DeepSeek-R1-0528 | May 2025 | Reasoning | 671B (37B) | 128K | Open weights | Fewer hallucinations, tool calls |
| DeepSeek-Prover-V2 | Apr 2025 | Math · theorem proving | 671B | — | Open weights | Formal proofs in Lean 4 |
| Janus-Pro | Jan 2025 | Image understanding + generation | 1B / 7B | — | Open weights | Unified multimodal model |
| DeepSeek-R1 + distills | Jan 2025 | Reasoning | 671B (37B) · distills 1.5B–70B | 128K | Open weights | The model that went viral |
| DeepSeek-V3 | Dec 2024 | General | 671B (37B) | 128K | Open weights | Low-cost MoE breakthrough |
| DeepSeek-VL2 | Dec 2024 | Vision-language | Tiny / Small / Base (MoE) | — | Open weights | Successor to DeepSeek-VL |
| DeepSeek-V2.5 | Sep 2024 | General + code | 236B (21B) | 128K | Open weights | Merged chat and coder models |
| DeepSeek-Coder-V2 | Jun 2024 | Code | 236B (21B) · Lite 16B | 128K | Open weights | 338 programming languages |
| DeepSeek-V2 | May 2024 | General | 236B (21B) | 128K | Open weights | Introduced MLA attention |
| DeepSeek-VL | Mar 2024 | Vision-language | 1.3B / 7B | — | Open weights | First vision model |
| DeepSeek-Math | Feb 2024 | Math | 7B | 4K | Open weights | Introduced GRPO training |
| DeepSeek-Coder | Nov 2023 | Code | 1.3B–33B | 16K | Open weights | First code models |
| DeepSeek LLM | Nov 2023 | General | 7B / 67B | 4K | Open weights | DeepSeek's first LLMs |
Sizes are total parameters, with active parameters per token in brackets for mixture-of-experts models. "Open weights" means weights are published on DeepSeek's Hugging Face page; check each model card for its license. "—" = not applicable or not published. Linked names open our detailed guides.
DeepSeek V4.1, V4, V3.2, V3.1, R1 & V3 Explained
What changed in each major DeepSeek release, from the V3 breakthrough to V4.1 Flash. Pick a model on the timeline to see its specs, key improvements and benchmark highlights.
DeepSeek V4.1 Flash
Released 10 Sep 2026The smallest model in DeepSeek's new architecture family, and the first with native visual understanding. DeepSeek calls it smarter, faster and more efficient, with results ahead of its own flagship V4-Pro.
- New asymmetric causal encoder–decoder architecture: reading input is cheaper than writing output.
- KV cache needs 1/4 the HBM and 1/8 the SSD of the previous generation, cutting agent costs.
- Native image understanding replaces the separate V4-Flash-Vision-Exp model.
- Open weights and technical report on Hugging Face.
- DeepSeek first planned to route V4-Pro traffic to it, then reversed that on 11 Sep 2026; V4-Pro stays available.
A new generation launched as an open-source preview in two sizes, Pro and Flash, making a 1M-token context the default across all DeepSeek services. GA builds followed (Flash 0731, Pro 0813).
- Token-wise compression plus DeepSeek Sparse Attention makes 1M-token context cost-effective.
- DeepSeek says V4-Pro led open models on agentic coding, maths, STEM and world knowledge at launch.
- Thinking and non-thinking modes; OpenAI and Anthropic API formats.
- Retired the old
deepseek-chat/deepseek-reasonernames on 24 Jul 2026. - Built to plug into agents such as Claude Code, OpenClaw and OpenCode.
The official successor to V3.2-Exp, which introduced DeepSeek Sparse Attention for cheaper long-context use. DeepSeek described V3.2 as GPT-5-level, with a high-compute Speciale variant.
- DeepSeek's first model to integrate thinking directly into tool use, in both thinking and non-thinking modes.
- Agent training data from 1,800+ environments and 85k+ complex instructions.
- V3.2-Speciale reached gold-level results at IMO, CMO, ICPC World Finals and IOI 2025 (API-only, briefly).
- Weights and tech report for both models on Hugging Face.
One model with two modes: Think and Non-Think, switched by the DeepThink button. It reached answers faster than R1-0528 and brought big gains in tool use and agent tasks.
- Hybrid inference:
deepseek-chat= non-thinking,deepseek-reasoner= thinking. - Stronger tool use and multi-step agent tasks after post-training.
- Added Anthropic API format support and strict function calling (beta).
- V3.1-Terminus (Sep 2025) fixed Chinese/English mixing and improved code and search agents.
The reasoning model that made DeepSeek famous, matching leading US models at a fraction of the reported cost. The R1-0528 update improved benchmarks, cut hallucinations and added JSON output and function calling.
- Thinks step by step before answering (chain of thought).
- R1-0528: reduced hallucinations and better front-end code generation.
- R1-0528 added JSON output and function calling.
- Its role is now covered by thinking mode in newer models; weights remain on Hugging Face.
The large, open mixture-of-experts model that proved frontier-level AI could be trained and served cheaply. It's the foundation that R1, V3.1 and V3.2 were built on.
- 671B MoE parameters with only 37B active per token.
- Trained on 14.8T high-quality tokens.
- 60 tokens/second, 3× faster than V2.
- Fully open-source models and paper.
- Launch API price: $0.27/M input (cache miss), $0.07/M (cache hit), $1.10/M output.
Facts from DeepSeek's official announcements; benchmark figures that those pages show only as images are taken from DeepSeek's API change log. All figures are DeepSeek's own and may differ from independent tests. Parameter counts for V3.1, V3.2 and R1 follow the V3 base they're built on.
How DeepSeek Models Work: MoE, Sparse Attention & More
The technology behind DeepSeek V4 and V4.1, and why it lets DeepSeek offer frontier-level AI at a fraction of the usual price.
V4-Pro has 1.6T total parameters but activates ~49B per token. V4.1 Flash has 552B total and activates only 8B to read input and 16B to generate output.
V4.1's new asymmetric design uses a smaller active path for reading context than for writing output, cutting the cost of the long inputs that dominate agent workloads.
V4.1 Flash's KV cache needs about 1/4 the HBM and 1/8 the SSD storage of the previous generation. Since cache hits are a big share of agent bills, this flows straight into lower prices.
Reduces long-context attention cost from O(L²) toward O(kL), making the 1M-token context window practical for whole-repository and book-length inputs.
V4.1 Flash understands images natively rather than via a separate vision model, replacing the retired V4-Flash-Vision-Exp. V4-Pro remains text-only.
Choose low, high or max reasoning effort per request. Thinking mode is on by default; non-thinking mode is available for fast, cheap responses.
DeepSeek Benchmarks: V4.1 Flash vs V4-Pro vs Claude Opus 4.8
How DeepSeek V4.1 Flash and V4-Pro compare with Claude Opus 4.8 and Kimi K3 on coding, agent and reasoning benchmarks, and on price.
● BENCHMARK SCORES
● FULL MODEL COMPARISON TABLE
| Metric | V4.1 Flash | V4-Pro | Opus 4.8 | Kimi K3 |
|---|---|---|---|---|
| Output price /1M | $0.60 (off-peak) | $1.98 (off-peak) | $25.00 | — |
| Context window | 1M | 1M | — | — |
| Vision input | ✓ Native | ✗ | — | — |
| Open weights | ✓ | ✓ | ✗ | — |
| Terminal-Bench 2.1 | 90.6 | 87.9 | 85.0 | 88.3 |
| HLE (no tools / tools) | 36.8 / 63.9 | 42.7 / 60.0 | 49.8 / 57.9 | — |
| NL2Repo | 65.4 | 61.5 | 69.7 | — |
| CyberGym | 88.1 | 83.3 | 78.3 | — |
| DeepSWE | 74.2* | 62.7 | — | 67.5 |
| Agents' Last Exam | 31.8 | 25.7 | — | — |
| SWE-bench Verified | — | 80.6 | 88.6 | — |
| Toolathlon-Verified | — | 74.1 | — | 76.5 |
| GPQA Diamond | 90.9 | — | — | — |
V4.1 Flash and V4-Pro scores are from DeepSeek's official change log (10 Sep and 13 Aug 2026); competitor scores from vendor model cards as of Aug–Sep 2026. *V4.1 Flash was tested on DeepSWE v1.1, so it is not directly comparable with the other DeepSWE figures. "—" = not reported on a comparable setting. Vendor-reported benchmarks often differ from independent tests. Deeper dives: DeepSeek vs Claude · DeepSeek vs GPT-5 · all models compared.
How to Use DeepSeek AI: Web, App & API
Three ways to use DeepSeek AI, step by step: chat free in your browser, use the mobile app, or connect your own software through the API.
Use it in your browser
- Go to chat.deepseek.com (DeepSeek's official site).
- Sign up with your email or a supported login.
- Type your question, or attach an image or file.
- Turn on deeper thinking for hard problems, then read and refine the answer.
Use the mobile app
- Open the App Store or Google Play and search "DeepSeek".
- On Android you can also get the APK from DeepSeek's official download page; avoid APKs from other sites.
- Log in with the same account to sync your chats.
- Ask by typing, speaking, or snapping a photo.
Build with the API
- Create an account at platform.deepseek.com and top up a small balance.
- Generate an API key and keep it secret.
- Point the OpenAI or Anthropic SDK at DeepSeek's base URL.
- Call
deepseek-flashordeepseek-v4-pro, ideally off-peak for half price.
Illustrations are simplified and don't show DeepSeek's actual interface. deepseeksr1.com is an independent guide; we don't host chat, apps or API keys.
DeepSeek Web: Free AI Chat in Your Browser ↗
DeepSeek Web is the official chat at chat.deepseek.com. It runs in any modern browser on a computer, tablet or phone, with nothing to install, and it's free with an account.
Chat in any browser
Chrome, Edge, Safari or Firefox, on desktop or mobile, with no download.
Deeper thinking
Switch on reasoning for maths, logic and coding; leave it off for quick answers.
Web search
Let it look up recent information and show sources, where the option is available.
Files & images
Upload PDFs, documents, screenshots or photos and ask questions about them.
Chat history
Past conversations stay in the sidebar and sync with the mobile app.
Free
No subscription. Usage can be slowed or paused when servers are busy.
Web, app or API: which should you use?
| Web | Mobile app | API | |
|---|---|---|---|
| Cost | Free | Free | Pay per token |
| Install needed | No | Yes | No (code) |
| Best for | Long work at a computer, files | Quick questions, voice, photos | Building apps & automation |
| Chat history sync | ✓ | ✓ | — |
| Where | chat.deepseek.com | Download | DeepSeek API |
Tips for DeepSeek Web
- One topic per chat. Start a new chat for each task; very long chats get slower and less focused.
- Bookmark it. Save chat.deepseek.com so you never land on a look-alike site from search ads.
- Make it feel like an app. In Chrome or Edge, use the browser menu to install the page as an app; on a phone, add it to your home screen.
- Paste, don't screenshot, text. Copy text in directly when you can, and save uploads for images and documents.
- Busy message? Wait a minute and resend; see DeepSeek Not Working?
- Mind privacy. Don't paste passwords or confidential data; chats are stored on DeepSeek's servers.
The illustration is simplified and doesn't reproduce DeepSeek's actual interface. Available buttons and features can change and may vary by region.
DeepSeek Login: How to Sign In Safely
How to sign in to DeepSeek safely on the web, in the app and on the API platform. You only ever log in on DeepSeek's own sites; deepseeksr1.com has no accounts and never asks for your password.
Web chat
- Type chat.deepseek.com into your browser yourself, or use a saved bookmark.
- Choose one of the sign-in options shown. These vary by region and may include email and password, a phone verification code, or a third-party account.
- Enter any verification code DeepSeek sends you.
Mobile app
- Open the official DeepSeek app (see Download).
- Sign in with the same method you used on the web so your chats sync.
API platform
- Developers sign in at platform.deepseek.com to manage API keys, top-ups and usage.
- Chatting is free; the API console is only needed for building apps.
Check the address before you type a password
The real login lives on deepseek.com addresses. Other sites using the DeepSeek name, with different endings or extra words, aren't run by DeepSeek. Don't enter your password there, even if they look identical. (Example addresses shown are made up.)
Login problems?
Forgot your password
- On the official login page, use the "forgot password" option if your sign-in method has one.
- Follow the reset link or code sent to your email or phone; it expires quickly.
- If you signed up with a third-party account, sign in with that instead; there may be no separate password.
Verification code doesn't arrive
- Check spam or promotions folders, and wait a minute before resending.
- Make sure the email or phone number has no typos.
- See DeepSeek Not Working? for more fixes.
"Account not found" on the API platform
- Make sure you're using the same sign-in method you originally signed up with.
- If you've never used the API, you may need to sign up on the platform first.
Keep your account safe
- Use a strong password you don't use anywhere else.
- Never share verification codes. DeepSeek staff won't ask for them.
- Log out on shared or public computers.
- Keep API keys secret, and delete any key you think has leaked.
DeepSeek Download: Android, iPhone, Windows & Mac
DeepSeek offers a free chat app for phones and a desktop app for Windows and Mac. Always download from DeepSeek's own download page or the official store listing; we don't host any files.
Android
DeepSeek chat app · free- Open deepseek.com/download on your phone and tap Android.
- If prompted, allow your browser to install apps, then open the downloaded file.
- Turn that permission back off after installing, then log in.
Only install an APK that comes from download.deepseek.com. Copies on other sites may be modified or contain malware.
Official Android download ↗iPhone & iPad
DeepSeek chat app · free- Open the App Store and search "DeepSeek", or use the link below.
- Check the developer is DeepSeek, then tap Get.
- Open the app and log in with the same account you use on the web.
Availability varies by country. If it isn't in your store, use chat.deepseek.com in Safari.
Open in App Store ↗Windows
DeepSeek Harness desktop- Go to deepseek.com/download and choose Windows.
- Run the downloaded
.exeinstaller and follow the prompts. - Open DeepSeek Harness and follow its first-run setup.
DeepSeek says the desktop app is still being refined. Just want to chat? Use the web chat in Edge or Chrome; both let you install a site as an app from their menu.
Official Windows download ↗Mac
DeepSeek Harness desktop- Go to deepseek.com/download and choose macOS.
- Open the
.dmgand drag the app into Applications. - Launch it from Applications and follow the setup.
Intel Macs aren't supported by the desktop app; use the web chat in Safari or Chrome instead.
Official Mac download ↗🛡️ Safe download checklist
DeepSeek Harness: Open-Source AI Agent Workspace ↗
DeepSeek Harness (DSH) is DeepSeek's open-source AI agent workspace. Point it at a folder and it can organise documents, analyse spreadsheets, write and test code, research with sources, and run tasks on a schedule, all extendable with plugins. It's MIT-licensed and runs as a desktop app or a local web UI.
Complete a range of tasks
Organise documents, analyse spreadsheets, write code. Works with Word, Excel, PDF, HTML, Markdown, Python and TypeScript files, with a review view showing diffs.
Everything is a plugin
Tools, skills and even the interface are plugins. Install them, or build one by chatting in Creator mode.
Adapt to your workflow
The Scheduled tasks plugin runs work on demand or on a schedule, such as a weekly report every Friday.
Developer tools
Execution traces and runtime details help you debug tool calls and see what the agent actually did.
Ways to run it
# start the web UI (requires Node.js)
npx @deepseek-ai/dsh webGet started in 4 steps
- Install the desktop app, or launch the web UI with the command shown.
- Open Settings → Models and add a DeepSeek API key (from platform.deepseek.com). Other OpenAI-compatible providers can also be configured.
- Click Choose workspace and pick the project folder the agent may work in.
- Start a session and describe the task. Approve file changes or commands when it asks.
Built-in & official plugins
More on GitHub under the dsh-plugin topic. Harness is built on the Cordis plugin framework.
Details from DeepSeek's Harness page and developer quickstart. Illustration is simplified and doesn't reproduce the real interface. See DeepSeek's Harness privacy policy and terms.
DeepSeek AI Use Cases for Developers, Business & Students
Pick who you are to see practical ways to use DeepSeek, which model fits each job, how much thinking effort to set, and a prompt you can copy straight into the official chat or the API.
Fix bugs across a whole repo
Paste the failing test output and the relevant files. Find the root cause, explain it in two sentences, then give a minimal patch as a unified diff.
Run a coding agent
Hook deepseek-flash into Codex, Claude Code or OpenCode via DeepSeek's documented integrations and let it plan, edit and test in your terminal.
Turn a screenshot into code
Here is a screenshot of a settings page. Rebuild it as a single responsive HTML + Tailwind component. Keep spacing and hierarchy, skip real data.
Write tests and docs
Write pytest unit tests for this module covering edge cases, then a short README section explaining how to use its public functions.
Customer-support assistant
You are a support agent for an online store. Using only the policy text below, answer the customer politely and say when to escalate to a human.
Read invoices and forms
Extract supplier, invoice number, date, line items and total from this invoice image. Return valid JSON matching the schema below.
Summarise long contracts
Summarise this contract for a non-lawyer: parties, term, payment, termination, liability caps and anything unusual. Flag clauses worth a lawyer's review.
Automate workflows
Given these tools (create_ticket, send_email, lookup_order), process each incoming message below and decide which tool calls to make.
Step-by-step tutoring
I'm stuck on this calculus problem. Don't give the answer yet: ask me one guiding question at a time and check my reasoning.
Solve a photographed problem
Here's a photo of my physics homework question. Explain which principle applies and walk through the solution with units.
Make a quiz
Create 10 multiple-choice questions on the French Revolution for 15-year-olds, with answers and a one-line explanation for each.
Plan a lesson
Draft a 45-minute lesson plan introducing photosynthesis, with a starter activity, main task, differentiation ideas and an exit question.
Review a stack of papers
Here are 12 abstracts. Group them by method, note where findings conflict, and list three open questions none of them answer.
Analyse a dataset
Here is a CSV sample and column descriptions. Write pandas code to clean it, find the three strongest trends, and plot each one.
Read charts and figures
Describe what this chart shows, extract the underlying numbers as a table, and point out anything misleading about the axes.
Hard maths and proofs
Prove the following statement rigorously. If it's false, give a counterexample. Show every step.
Brainstorm ideas
Give me 15 YouTube video ideas about budget travel in Sri Lanka, each with a hook title and a one-line angle.
Edit a long manuscript
Here is my 60,000-word draft. Flag plot holes, inconsistent character details and pacing problems, chapter by chapter.
Translate and localise
Translate this product description into Sinhala and Tamil, keeping the brand tone friendly and adapting idioms for local readers.
Repurpose content
Turn this blog post into a LinkedIn post, five tweets and a 60-second video script, each in a matching tone.
Don't paste confidential, personal or regulated data into the hosted service; see Why Select DeepSeek AI? for when to self-host instead.
Best DeepSeek Prompts and a Free Prompt Builder
Better prompts get better answers. Learn the five parts of a strong prompt, build your own in seconds, or copy one of our ready-made prompts into the official DeepSeek chat.
Role
Who should DeepSeek act as?
"Act as a senior tax adviser…"Task
What exactly do you want done?
"…explain the new VAT rules…"Context
Background it can't guess.
"…for a small café in Colombo…"Format
How the answer should look.
"…as 5 bullet points…"Constraints
Limits and things to avoid.
"…under 150 words, no jargon.""Write about marketing."
DeepSeek has to guess the audience, length, angle and format, so you get a generic essay."Act as a marketing consultant. Suggest 5 low-budget Instagram ideas for a new bakery in Kandy aimed at university students. Give each a one-line hook and an estimated cost. Keep it under 200 words."
Role, task, context, format and constraints are all there, so the answer is usable straight away.Prompt builder
Your prompt
Tips for DeepSeek
- Turn on deeper thinking (or set effort to high or max on the API) for maths, logic and coding; keep it off for quick chats.
- V4.1 Flash can read images, so attach screenshots, charts or photos instead of describing them.
- On the API, keep long fixed instructions at the start of every request so they're cached and billed cheaply.
- If the answer misses, reply with what to change ("shorter", "add costs") rather than starting over.
Prompt library
Blog post outline
Act as an experienced content editor. Create a detailed outline for a 1,200-word blog post titled "[TITLE]" for [AUDIENCE]. Include an intro hook, 5–7 H2 sections with bullet points, and a conclusion with a call to action.
Rewrite in a clearer tone
Rewrite the text below so it is clear, friendly and easy to read for a general audience. Keep every fact, cut filler, and keep it under [N] words.
[PASTE TEXT]
Professional email reply
Draft a polite, concise reply to the email below. My goal is to [GOAL]. Offer two subject-line options and keep the body under 150 words.
[PASTE EMAIL]
Explain this code
Explain what the code below does, line by line, for a junior developer. Then list any bugs, edge cases or security issues you see, most serious first.
[PASTE CODE]
Write a function with tests
Write a [LANGUAGE] function that [WHAT IT SHOULD DO]. Include type hints, a docstring, input validation, and unit tests covering normal cases and edge cases.
Debug an error
I get this error when running my program. Explain the likely cause, show how to confirm it, and give the smallest fix.
Error:
[PASTE ERROR]
Code:
[PASTE CODE]
Explain like I'm new
Explain [TOPIC] to a complete beginner in under 200 words, using one everyday analogy. Then give three quick questions to check my understanding.
Make flashcards
Turn the notes below into 15 question-and-answer flashcards. Keep answers to one sentence and focus on facts likely to appear in an exam.
[PASTE NOTES]
Check my working
Here is my solution to a maths problem. Don't redo it from scratch: find the first step where I went wrong, explain why, and show the correct step.
[PASTE PROBLEM AND SOLUTION]
SWOT analysis
Create a SWOT analysis for [BUSINESS] in [MARKET/LOCATION]. Give four points per quadrant, each with a one-line reason, then the two most important actions to take next.
Meeting notes to actions
Turn the meeting transcript below into: a 3-sentence summary, a table of action items (owner, task, deadline), and open questions.
[PASTE TRANSCRIPT]
Extract data as JSON
Extract every [ITEM TYPE] from the text below. Return only valid JSON: an array of objects with the fields [FIELD 1], [FIELD 2], [FIELD 3]. Use null for anything missing.
[PASTE TEXT]
Compare two options
Compare [OPTION A] and [OPTION B] for [USE CASE]. Make a table covering cost, ease of use, performance and risks, then recommend one and explain when the other would be better.
Describe a chart
Look at this chart image. Describe what it shows in plain language, extract the data as a table, and point out any misleading scales or missing labels. (Attach the image; works with V4.1 Flash.)
Screenshot to steps
This is a screenshot of an app screen. Write numbered steps telling a new user how to [TASK] from this screen. (Attach the screenshot; works with V4.1 Flash.)
No prompts match your search.
Replace the [BRACKETED] parts with your own details. More techniques in our prompting guide. Never paste passwords, ID numbers or confidential data into any AI chat.
DeepSeek API: Endpoints, Features & Quick Start
DeepSeek's API lets your own apps, scripts and agents call its models. It speaks both OpenAI and Anthropic formats, so most existing code works after changing two settings.
Key details
- Base URL (OpenAI format)
https://api.deepseek.com- Base URL (Anthropic format)
https://api.deepseek.com/anthropic- Models
deepseek-flash(V4.1 Flash) ·deepseek-v4-pro- Context / max output
- 1M tokens / 384K tokens
- Concurrency limit
- 2,500 (Flash) · 500 (V4-Pro)
- Billing
- Prepaid balance, per token, off-peak 50% off
- Old names
deepseek-v4-flash,-vision-exp→ temporary aliases to Flash
What the API can do
Chat Completions
OpenAI-compatible /chat/completions, with streaming.
Responses API
OpenAI Responses format, adapted for Codex-style agents.
Anthropic format
Use Anthropic SDKs and tools via /anthropic.
Tool calls
Function calling for agents and workflow automation.
JSON output
Force valid JSON for extraction and structured data.
Vision
Send images to deepseek-flash (not V4-Pro).
Thinking mode
On by default; switch off or set low / high / max effort.
Automatic caching
Repeated input is cached and billed ~98% cheaper on Flash.
FIM & prefix (beta)
Fill-in-the-middle and chat-prefix completion, non-thinking mode only.
Quick start: OpenAI-format code
from openai import OpenAI
# DeepSeek uses an OpenAI-compatible API
client = OpenAI(
api_key="<your-deepseek-api-key>",
base_url="https://api.deepseek.com"
)
response = client.chat.completions.create(
model="deepseek-flash", # V4.1 Flash, or deepseek-v4-pro
messages=[
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": "Explain MoE architecture"}
],
stream=False
)
print(response.choices[0].message.content)
import OpenAI from 'openai';
const client = new OpenAI({
apiKey: '<your-deepseek-api-key>',
baseURL: 'https://api.deepseek.com',
});
const completion = await client.chat.completions.create({
model: 'deepseek-flash',
messages: [
{ role: 'user', content: 'Hello!' }
],
});
console.log(completion.choices[0].message.content);
curl https://api.deepseek.com/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer YOUR_KEY" \
-d '{
"model": "deepseek-flash",
"messages": [
{"role":"user","content":"Hello!"}
]
}'
Already using the Anthropic SDK?
Point it at DeepSeek's Anthropic-format endpoint:
# pip install anthropic
import anthropic
client = anthropic.Anthropic(
api_key="<your-deepseek-api-key>",
base_url="https://api.deepseek.com/anthropic",
)
msg = client.messages.create(
model="deepseek-flash",
max_tokens=1024,
messages=[{"role": "user", "content": "Hello!"}],
)
print(msg.content[0].text)Works with popular coding agents
DeepSeek's docs include setup guides for these tools:
Good practice
- Keep your API key on the server, never in browser or app code.
- Put stable instructions first so they get cached and billed cheaply.
- Run batch jobs off-peak for half price.
- Call
deepseek-flash, not the old aliases, which are temporary. - Handle rate-limit errors with retries and back-off.
Details from DeepSeek's official Models & Pricing page and change log. Our API guide goes deeper.
DeepSeek Pricing: API Costs and Token Calculator
Current DeepSeek API pricing for V4.1 Flash and V4-Pro, how tokens are billed, and a calculator showing what your own usage will cost. The web chat and apps are free.
Official API prices
| Model (API name) | Context | Input (Cache Hit) | Input (Cache Miss) | Output | Status |
|---|---|---|---|---|---|
| V4.1 Flash (deepseek-flash) | 1M / 384K out | $0.003 ($0.006 peak) | $0.15 ($0.30 peak) | $0.60 ($1.20 peak) | Current · vision · 2,500 concurrency |
| V4-Pro-0813 (deepseek-v4-pro) | 1M / 384K out | $0.022 ($0.044 peak) | $0.66 ($1.32 peak) | $1.98 ($3.96 peak) | Current · text only · 500 concurrency |
| deepseek-v4-flash / deepseek-v4-flash-vision-exp | 1M | Retired 10 Sep 2026 · served by V4.1 Flash at Flash prices | Temporary alias | ||
| deepseek-chat / deepseek-reasoner | 128K | Retired 24 Jul 2026 · weights still free to self-host | Retired | ||
USD per 1M tokens, off-peak (peak in brackets), from DeepSeek's official Models & Pricing page. Flash prices effective 04:00 UTC, 10 Sep 2026. DeepSeek can change prices at any time, so always check the official page before budgeting.
How token costs work
What is a token?
A token is a small chunk of text, often a word or part of a word. In English, 1,000 tokens is roughly 750 words. You pay for the tokens you send (input) and the tokens the model writes back (output).
- Cache hit: input DeepSeek has seen recently, such as a repeated system prompt, billed at about 2% of the normal input price on Flash.
- Cache miss: new input, billed at the standard input rate.
- Output: everything the model writes, including its thinking when thinking mode is on.
- Peak hours (01:00–04:00 and 06:00–10:00 UTC, weekdays) cost 2×. Everything else is off-peak.
Token cost calculator
Uses DeepSeek's published USD list prices; actual bills can differ. Check the official pricing page before budgeting.
What everyday tasks cost (off-peak)
| Task | Quantity | V4.1 Flash | V4-Pro |
|---|---|---|---|
| 💬One chat message ~500 tokens in · ~300 out | per 1,000 messages | $0.255 | $0.924 |
| 📄Summarise a 20-page PDF ~12K in · ~800 out | per 100 documents | $0.228 | $0.950 |
| 🖼️Read an invoice image ~2K in · ~400 out | per 1,000 invoices | $0.540 | $2.11 |
| 🤖One coding-agent session ~2M in (90% cached) · ~40K out | per session | $0.059 | $0.251 |
| 📚Analyse a whole book ~600K in · ~4K out | per book | $0.092 | $0.404 |
Token counts are rough estimates; real usage depends on language, file content, and how much the model thinks. Peak-hour costs are double. The free official chat and apps cost nothing; these prices apply only to the API.
DeepSeek Not Working? How to Fix Common Problems
"Server is busy", login trouble, app crashes or API errors? Start with a quick status check, then pick where you're having the problem for step-by-step fixes.
Quick 4-step check
- Is it down for everyone? Check DeepSeek's official status page.
- Wait and retry. "Server busy" usually clears within minutes.
- Refresh or restart the page or app, and check your internet.
- Try another route: web instead of app, or a different browser or network.
"Server is busy. Please try again later."
- This is DeepSeek's capacity limit, not a problem with your device. Wait 30–60 seconds and send again.
- Check status.deepseek.com for an outage.
- Busy spells are worst right after big launches and in Asian daytime hours, so try again later.
- If you need answers now, shorten the request or start a fresh chat.
The page won't load or stays blank
- Hard-refresh the page (Ctrl/Cmd + Shift + R).
- Clear cookies and cache for chat.deepseek.com, or try a private window.
- Turn off ad blockers, VPNs or privacy extensions one at a time.
- Try a different browser or network (e.g. switch Wi-Fi to mobile data).
I can't log in or the verification code never arrives
- Check your spam or promotions folder.
- Wait a minute before tapping "resend"; repeated requests can be throttled.
- Double-check the email address or phone number for typos.
- Try another sign-in option offered on the login page.
The answer stops halfway or is cut off
- Type "continue" to pick up where it stopped.
- Very long chats get heavy, so start a new chat and paste a short summary.
- Split huge documents into smaller parts.
- Regenerate the response if it looks broken.
File or image upload fails
- Check the file type and size; try a smaller or compressed version.
- Paste the text directly instead of uploading a scanned PDF.
- Re-upload after refreshing the page.
I can't find the app in my store
- DeepSeek's app may not be offered in every country or on every store.
- Search for the app from developer "DeepSeek" only; avoid look-alike apps.
- Android users can get the APK from DeepSeek's official download page; never from other websites.
- Use the web chat in your phone's browser instead.
The app crashes or won't open
- Update the app and your phone's operating system.
- Restart your phone.
- Android: Settings → Apps → DeepSeek → Clear cache.
- Still crashing? Uninstall and reinstall, then log in again.
Chats aren't syncing between devices
- Make sure you're logged in to the same account everywhere.
- Pull down to refresh the chat list.
- Check your internet connection, then log out and back in.
Voice input or camera doesn't work
- Allow microphone and camera permission for DeepSeek in your phone's settings.
- Close other apps using the mic or camera.
- Restart the app after changing permissions.
| Code | Meaning | Fix |
|---|---|---|
| 400 | Invalid Format Request body is malformed | Fix the body using the hints in the error message. |
| 401 | Authentication Fails Wrong API key | Check the key, or create a new one on the platform. |
| 402 | Insufficient Balance Account has run out of credit | Top up your balance. |
| 422 | Invalid Parameters A parameter isn't valid | Correct it using the error message hints. |
| 429 | Rate Limit Reached Too many concurrent requests | Slow down and retry with back-off; limits are 2,500 (Flash) / 500 (Pro) per account. |
| 500 | Server Error Problem on DeepSeek's side | Wait briefly and retry; contact DeepSeek if it persists. |
| 503 | Server Overloaded High traffic | Wait briefly and retry. |
Requests hang for a long time
- DeepSeek keeps slow connections open instead of failing them: non-streaming calls receive empty lines and streaming calls receive ": keep-alive" comments while waiting.
- Set a generous client timeout; DeepSeek closes the connection only if inference hasn't started after 10 minutes.
- If you parse raw HTTP yourself, ignore those keep-alive lines.
- Use streaming so users see output as soon as it starts.
"Model not found" or unexpected model
- Use
deepseek-flashordeepseek-v4-pro. deepseek-chatanddeepseek-reasonerwere retired on 24 Jul 2026.deepseek-v4-flashstill works only as a temporary alias, so switch now.
Images are ignored
- Only
deepseek-flash(V4.1 Flash) supports vision; V4-Pro is text-only.
Bills are higher than expected
- Peak hours (01:00–04:00 and 06:00–10:00 UTC, weekdays) cost double.
- Thinking tokens count as output, so lower the effort for simple tasks.
- Keep fixed instructions at the start of each request to get cache-hit pricing.
DeepSeek News: Latest Updates
The latest DeepSeek news: new models, API changes, pricing and company updates, checked against DeepSeek's official change log.
DeepSeek-V4.1-Flash: 552B open-weight MoE with native vision, and V4-Pro isn't going anywhere yet
DeepSeek's release notes and technical report fill in the specs for V4.1 Flash: a 552B-parameter mixture-of-experts with a new causal encoder–decoder design that activates just 8B parameters to read input and 16B to write output. It understands images natively, keeps the 1M-token context, and its KV cache needs about 1/4 of the HBM and 1/8 of the SSD of the previous generation. Weights and the technical report are on Hugging Face.
On the API it is called deepseek-flash; V4-Flash and V4-Flash-Vision-Exp are retired and their old names temporarily route to V4.1 Flash. DeepSeek originally planned to route all V4-Pro traffic to V4.1 Flash from 14 September, but reversed that on 11 September after user pushback: deepseek-v4-pro keeps running at its existing price until further notice.
Benchmarks are DeepSeek's own published figures. Third-party analysis of the model card notes V4.1 Flash leads V4-Pro on agentic tests but trails it on most knowledge-heavy base-model tests.
Bloomberg and Reuters report DeepSeek is close to raising at least 80 billion yuan, with Tencent and CATL among the largest backers, well above its original ~50 billion yuan target. Sources tie the extra demand to the V4.1 Flash launch. DeepSeek has not announced the round; an IPO on Shanghai's STAR Market is reportedly being prepared for early 2027.
A Bloomberg Intelligence report says top Chinese models now trail US rivals by about 3% on benchmark scores after V4.1 Flash, down from roughly 9% in May.
Reuters reported DeepSeek would join OpenAI, Anthropic and other labs in briefing the Security Council on AI risks. DeepSeek also announced work with Huawei on programming tools for Ascend AI chips.
"In response to user demand," DeepSeek said it will keep serving V4-Pro after 14 September with billing unchanged, instead of routing it to V4.1 Flash. A V4.1 Pro is confirmed to be in development, with no date given.
New Flash prices took effect at 04:00 UTC. The same day, China's Commerce Ministry rejected the US advisory (below), calling distillation a common industry practice and warning of countermeasures if it is used to target Chinese firms.
An NSA, FBI and CISA advisory named DeepSeek, Alibaba, Moonshot AI, MiniMax, StepFun and Z.AI, alleging they extract US models in breach of terms of use.
DeepSeek FAQ: Frequently Asked Questions
Quick answers to the most common questions about DeepSeek AI.
DeepSeek-V4.1-Flash, released 10 September 2026, is the smallest model in DeepSeek's new architecture family. It is a 552B-parameter mixture-of-experts with a causal encoder–decoder design (8B active parameters for input, 16B for output), native image understanding, a 1M-token context and up to 384K output. Its KV cache needs about 1/4 the HBM and 1/8 the SSD of the previous generation. On the API it is called deepseek-flash, and the weights are on Hugging Face.
Not for now. DeepSeek initially announced that all deepseek-v4-pro requests would be routed to V4.1 Flash from 14 September 2026. On 11 September it reversed that, saying that in response to user demand it would keep serving V4-Pro with unchanged billing and give notice of any future change. A V4.1 Pro is in development, but no date has been announced.
Start with V4.1 Flash. It is cheaper (about 70% less per output token), supports images, has higher concurrency limits, and DeepSeek reports it ahead of V4-Pro on agentic coding benchmarks. Consider V4-Pro for knowledge-heavy work where it still scores higher, such as HLE without tools (42.7 vs 36.8).
Yes. DeepSeek's official web chat at chat.deepseek.com and its official mobile apps are free. The API is pay-per-token: off-peak, deepseek-flash costs $0.15 per million input tokens ($0.003 with cache hits) and $0.60 per million output tokens. See our DeepSeek Web guide.
Both were retired on 10 September 2026. For compatibility, the names deepseek-v4-flash and deepseek-v4-flash-vision-exp temporarily route to V4.1 Flash and are billed at Flash prices. DeepSeek describes this as temporary, so update your code to deepseek-flash.
For non-sensitive work the hosted API is widely used. However, DeepSeek is a Chinese company, data sent to its API may be stored on servers subject to Chinese law, and a September 2026 US security advisory accused it and other Chinese labs of unauthorised model distillation. For healthcare, finance, legal or government data, self-host the open weights or use a third-party cloud host with appropriate compliance and data-residency guarantees.
Yes, with the right hardware. On a laptop or single GPU, use Ollama or LM Studio with a distilled or quantized model (1.5B–70B, 8–48GB VRAM). The full V4-Flash (284B) fits a single 8-GPU server; V4.1 Flash (552B) and V4-Pro (1.6T) need multi-GPU clusters. See our self-hosting guide.
DeepSeek has confirmed that V4.1 Pro is in development but has given no release date, specs or pricing. We'll update this page as soon as DeepSeek's official change log confirms a release.
No. deepseeksr1.com is an independent guide and is not affiliated with DeepSeek. We don't offer chat, accounts, API keys or app downloads. Use deepseek.com, chat.deepseek.com and platform.deepseek.com for DeepSeek's own services.