What is DeepSeek and how it can help you be more productive

DeepSeek assistant guide · Updated August 2026

DeepSeek isn’t just “the Chinese ChatGPT” anymore. What started in January 2025 as a surprise open-weight release that shook global markets has become a permanent fixture of the AI landscape: a free chat app, an open model family you can download and run yourself, and API pricing that keeps dragging the entire industry down. In April 2026 DeepSeek shipped its V4 generation – V4 Pro and V4 Flash, both with a 1 million-token context window – and by August 2026 V4 Pro had exited preview with frontier-adjacent coding and reasoning scores at a fraction of Western rivals’ prices.

In this guide we explain what DeepSeek is in 2026, how the V4 Pro and V4 Flash models work, why open weights and MIT licensing matter, why R2 still hasn’t shipped, what the app and API cost, and how to use DeepSeek well.

Jan 2025R1 launches the “Sputnik moment” for open-weight AI
V4Current family: Pro · Flash · 1M context
MIT open weightsDownload, fine-tune, and self-host every model
01 · What it is

What is DeepSeek?

DeepSeek is a Chinese AI research lab – and the assistant built around its models – founded in 2023 by Liang Wenfeng, a quant fund manager who runs the hedge fund High-Flyer. The lab started as High-Flyer’s research arm before spinning out on its own. The name isn’t an acronym; it translates roughly to “deep exploration,” the company’s chosen nod to the idea of searching hard for better answers. DeepSeek caught the world’s attention in January 2025 when its R1 reasoning model matched Western frontier models at a fraction of the cost – triggering what markets dubbed a “Sputnik moment” for open-weight AI.

Today DeepSeek is a complete package, not just a research lab. Depending on where you use it, it can:

  • Answer questions & explain concepts
  • Write, rewrite, summarize & translate
  • Search the web with citations
  • Reason step by step with Think modes
  • Process 1M-token documents
  • Write, debug & review code
  • Analyze spreadsheets & data
  • Run locally on your own hardware
  • Serve cheap API calls at scale
DeepSeek – the models

DeepSeek is best known for its open-weight model family: V3 for general chat, R1 for step-by-step reasoning, and the current V4 generation – V4 Pro and V4 Flash – which folds reasoning into every response. Every model ships with weights you can download under the permissive MIT license.

DeepSeek – the app, the API, and the weights

Consumers meet DeepSeek through its free chat app (web and mobile). Developers meet it through the DeepSeek API – with OpenAI- and Anthropic-compatible endpoints – or through the model weights on Hugging Face, which anyone can self-host, fine-tune, or audit. The same models power all three surfaces.

02 · The model family

DeepSeek V4: Pro and Flash

DeepSeek’s current generation is V4, released as a preview on April 24, 2026 – exactly a year after the V3 era that rattled Silicon Valley. V4 ships in two sizes, both with a 1 million-token context window, both open-weight under the MIT license. V4 Pro left preview on August 13, 2026; V4 Flash remains the fast, low-cost workhorse.

Flagship reasoning · open weights

V4 Pro

The large V4 model: 1.6 trillion total parameters with about 49 billion active per token, trained on over 32 trillion tokens. It reports 80.6% on SWE-bench Verified and leads the open-weight field on coding and math, while staying roughly 8–15× cheaper than GPT-5.5 through the hosted API.

Maximum capabilityPreview ended Aug 2026
Fast & cheap · everyday default

V4 Flash

The smaller, faster sibling: 284 billion total parameters with about 13 billion active. It handles everyday chat, summarization, and high-volume API workloads at $0.14 per million input tokens – and a quantized copy can run on high-end consumer hardware.

Best cost/capability balance

Two notes on the wider lineup. First, the R series: R1 made DeepSeek famous in January 2025, and its successor R2 has been rumored, teased, and expected ever since – but as of August 2026 it still hasn’t shipped. DeepSeek appears to be holding it back rather than release something below its own bar, so R1-0528 remains the standalone reasoning option while V4 integrates reasoning directly into every response. Second, the older V3 family (V3, V3.1, V3.2) is still widely deployed and fine-tuned in the open-weight community, even though new development has moved to V4.

03 · How it works

How does DeepSeek work?

DeepSeek models are Mixture-of-Experts transformers: only a fraction of the total parameters are active for any given token, which is how a 1.6-trillion-parameter model can answer so cheaply. V4 adds a hybrid attention design – compressed sparse attention combined with a second attention path – that DeepSeek says cuts compute to about 27% of V3.2’s and shrinks the KV cache to a tenth at 1M-token context lengths.

V4 also folded reasoning into the main models instead of keeping it in a separate R-series line. You pick a thinking level per request:

  • 1

    Non-thinking mode

    The fastest, cheapest mode for routine questions, summarization, and drafting – answers in a single pass.

  • 2

    Think mode

    The default reasoning level for hard questions – the model works through the problem step by step before answering, like the old R1 line.

  • 3

    Think Max

    The deepest reasoning budget for the hardest problems – note that it needs a 384K-token minimum context window, so it consumes part of your context budget.

  • 4

    Tools on top

    Like its rivals, V4 combines the model with tools: live web search, file uploads, and the API surface that powers agentic coding workflows.

04 · What you can do

What can you do with DeepSeek?

Everyday chat

The free app handles questions, explanations, drafting, and translation. For most everyday tasks the difference between DeepSeek and a paid rival is barely noticeable – which is exactly why it shook the market.

Coding and debugging

V4 Pro’s vendor-reported 80.6% on SWE-bench Verified ties the top of the open-weight field. Developers use it through the API in coding agents, or run it locally for air-gapped or regulated projects.

Long-document analysis

The 1 million-token context window holds an entire codebase, a book-length document, or hours of transcripts in a single prompt – with up to 384,000 tokens of output for the answer.

Data analysis

Upload spreadsheets, CSVs, and PDFs; DeepSeek cleans the data, computes metrics, and explains the numbers rather than just pasting them back.

Self-hosting

The MIT license means the weights are yours. A quantized V4 Flash fits on high-end consumer machines with enough VRAM; V4 Pro realistically needs a multi-GPU server. Self-hosting keeps your data off any third-party servers.

Research with sources

Web search grounds answers in live results with citations instead of only training data – useful for current events, prices, and anything that moved after the model’s knowledge cutoff.

Writing and translating

Drafting, rewriting, and translation between languages are among its strongest everyday skills – especially Chinese-English, where it starts with a native understanding of both.

Learning and education

Free to use, strong at step-by-step explanations, and willing to walk through your mistakes – the same combination that made the app a global phenomenon in January 2025.

05 · Open weights

Why open weights matter

The single most important thing to understand about DeepSeek is the license. Every DeepSeek model – V3, R1, and the V4 family – ships under the permissive MIT license with weights published on Hugging Face. Anyone can download the full model, fine-tune it, integrate it, resell it, or run it on their own hardware with no payment and no approval. That is very different from “open” in the marketing sense: GPT-5.5, Claude, and Gemini are closed; Llama and Qwen are open-weight but with usage restrictions; DeepSeek is about as open as frontier models get.

01

Price pressure on the whole industry

When a lab can ship frontier-adjacent performance at $0.14 per million input tokens – or for free if you self-host – everyone else’s pricing has to move. Independent reviewers estimate V4 Pro is 8–15× cheaper than GPT-5.5 for similar workloads, and Western labs have repeatedly cut prices after DeepSeek releases.

The “Sputnik moment” that hit global markets in January 2025

02

Auditability and control

Open weights let security teams audit what the model actually does, let companies fine-tune on their own data, and let regulated industries run everything behind their own firewall. For many enterprises that outweighs a small capability gap.

MIT license · no usage restrictions

03

The China context

DeepSeek sits in the middle of the US-China AI competition: it built part of its training pipeline around Huawei Ascend chips after US export controls cut it off from top-end NVIDIA hardware, and US agencies have debated restricting access to its models. Use it for the model it is, and be aware governments on both sides are still deciding what it becomes.

A geopolitical story as much as a technology story

06 · Key moments

DeepSeek timeline: 2023–2026

2023

High-Flyer’s research arm becomes DeepSeek, a full AI lab; early DeepSeek Coder and Chat models arrive later that year.

Dec 2024

DeepSeek V3 launches: a 671-billion-parameter MoE model trained at a fraction of Western costs, with open weights that immediately feed the open-source community.

Jan 2025

R1, a step-by-step reasoning model, becomes a global phenomenon: the free app tops app-store charts, and markets briefly wipe roughly a trillion dollars off US tech stocks – the “Sputnik moment.”

Aug–Sep 2025

V3.1 and then V3.2 refine the general model; R1-0528 updates the reasoning line. R2 is rumored but never arrives.

Apr 2026

V4 Preview ships quietly – no press release, just model cards and API docs. V4 Pro (1.6T/49B active) and V4 Flash (284B/13B active), both with 1M context, MIT weights, and a 75% introductory price cut.

May 2026

DeepSeek makes the 75% discount permanent; the old deepseek-chat and deepseek-reasoner API ids are retired in July after a long deprecation window.

Aug 2026

V4 Pro exits preview (Aug 13) and API pricing moves to peak/off-peak windows from Aug 16. R2 still hasn’t shipped – the most-requested model in AI that doesn’t exist yet.

07 · Staying productive

How to use DeepSeek well

Be specific

State the audience, the goal, the format, and the constraints up front. “Summarize this into 5 bullet points for a busy founder” beats “summarize this.”

Pick the right model and mode

Everyday questions: V4 Flash, non-thinking. Hard reasoning or coding: V4 Pro with Think. Deepest problems: Think Max – and remember it consumes 384K of context budget.

Use the 1M window

Don’t summarize before asking. Paste the whole document, the whole codebase, or the whole transcript and ask the question directly – the context window is the point.

The prompt pattern that works

Context:Here is our full Q2 report (attached, 40 pages).Task:Summarize the quarter and flag the three biggest risks for the board.Constraints:One page. Bold key numbers. Risks as bullets with one line of evidence each.

Context, task, constraints – with a 1M window, give it everything and let it decide what matters. For sensitive or regulated work, self-host the weights and keep the data in-house.

08 · The window

How large is the context window?

Both V4 models support a standard 1 million-token context window – roughly 750,000 words – with up to 384,000 tokens of output. That is enough to hold an entire codebase or a book-length document in a single prompt, and the hybrid attention design is what keeps that long a window affordable to serve.

V4 Pro / V4 Flash – 1,000,000 tokens≈ 750,000 words
V3.2 / V3 / R1 (2024–2025) – 128,000 tokens≈ 96,000 words

Practical rule of thumb: a token is roughly 0.75 English words, and about 1.5 tokens per word once you factor in punctuation and spacing. Two caveats at the long end: attention is not uniformly strong across a 1M window – every frontier model shows some “lost in the middle” degradation – and Think Max mode requires a 384K-token minimum context, which eats into the budget your actual query would otherwise use.

09 · Plans and pricing

What does DeepSeek cost?

PlanPriceWhat you getLimits
DeepSeek app$0Free chat with V4 Pro / V4 Flash, web search, file uploadsNo consumer paid plan exists as of August 2026 – the app is free
API – V4 Pro$0.435 in / $0.87 out per 1MFlagship reasoning, 1M context, Think modes$0.003625 cached input; peak/off-peak pricing from Aug 16, 2026
API – V4 Flash$0.14 in / $0.28 out per 1MFast everyday tasks and high-volume workloads$0.0028 cached input
Self-hostedFree (MIT)Full model weights, unlimited use, data stays in-houseHardware is on you: quantized Flash fits high-end consumer gear; Pro needs a multi-GPU server

For context: V4 Pro launched in April 2026 at $1.74/$3.48 per million tokens with a 75% discount, which DeepSeek made permanent in May – the $0.435/$0.87 rates above are the current list prices. From August 16, 2026 the API moves to 2× peak-hour pricing during two daily windows around Beijing time. As with every AI vendor, prices and limits change often; the chart above reflects August 2026 and DeepSeek’s pricing page is the source of truth.

10 · FAQ

Frequently asked questions

Is DeepSeek free?

The consumer app is completely free – there is no paid consumer plan as of August 2026. Developers pay for API access (from $0.14 per million input tokens for V4 Flash), and the model weights themselves are free under the MIT license if you self-host.

Is DeepSeek better than ChatGPT?

On coding and math, V4 Pro is right at the top of the open-weight field and close behind the Western frontier – US government testing puts it roughly eight months behind American leaders overall. ChatGPT still wins on multimodal features, agentic execution, and ecosystem integrations, and DeepSeek V4 launched without native image or video support. Where DeepSeek wins decisively is value: comparable answers at a fraction of the cost, or free if you self-host.

Is DeepSeek open source?

The weights are open under the permissive MIT license – anyone can download, fine-tune, and sell them. Strictly speaking that makes DeepSeek “open weight” rather than fully open source, because the training data and training code aren’t published.

Is my data safe with DeepSeek?

DeepSeek is a Chinese company, and app data is processed in China, which has drawn privacy scrutiny from regulators and app stores around the world. For anything sensitive, professional, or regulated, either treat it as confidential data you shouldn’t upload – or self-host the open weights and keep everything on your own infrastructure.

Why hasn’t R2 shipped?

R2 has been the most-requested model that doesn’t exist: rumored and teased since 2025, it still hasn’t been released as of August 2026. DeepSeek appears to be holding it back rather than ship something below its own bar. In the meantime, V4 folds reasoning into every response, and R1-0528 remains available as a standalone reasoning model.

What are DeepSeek’s limitations?

V4 launched text-only with no native image, audio, or video generation. It can still hallucinate on obscure or fast-moving facts, long-context accuracy degrades in the middle of a 1M window, and the hosted API routes through servers outside most Western jurisdictions. US export-control debates also mean the model’s availability and regulation could change.

Heads-up:DeepSeek moves fast – models, API ids, and prices change frequently, and the details above reflect August 2026. Check DeepSeek’s official API docs and pricing page for current availability and rates, and treat any claimed R2 timeline as rumor until it actually ships.

Related reading