What Is DeepSeek? Models, Architecture and How to Use It
I've been using DeepSeek since the R1 moment in early 2025, and it's still the AI tool people ask me about most. What is DeepSeek? Short answer first: one name, two things — an AI research company in Hangzhou, and the free chat assistant and API that company ships. The models sit in the top tier, the prices are a fraction of everyone else's, and most of the weights can be downloaded and run on your own machines, no strings attached. The long answer — every model since 2023, the architecture that made it cheap, what the benchmarks actually say, and how to put it to work today — is what this article is for.
What Is DeepSeek?
The company: Hangzhou DeepSeek, full name Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd., registered July 17, 2023. Founder Liang Wenfeng, Zhejiang University degrees in electronic engineering and computer science, funded by High-Flyer, the quantitative fund he co-founded. The team runs about 160 people — unusually small for what it ships. That's the company level; the interesting part is the product.
The product line, as the official site footer lists it:
- DeepSeek web and app — the chat assistant, free, at chat.deepseek.com and on iOS/Android
- API platform — pay-as-you-go access to the same models, compatible with both OpenAI and Anthropic formats
- DeepSeek Harness — the official agent tooling, in developer preview
- API docs — documentation for all of the above
So when someone says "DeepSeek" they might mean the lab, the chatbot, the API, or the model weights on Hugging Face. Four doors, one building.
Every Model Since 2023, in Order
The lineup didn't arrive in one piece — a dozen-plus releases over three years, each adding something concrete:
| When | Model | What it added |
|---|---|---|
| Nov 2023 | DeepSeek Coder | First model, code-first (87% of the training corpus was code) |
| May 2024 | DeepSeek-V2 | First MoE architecture + MLA attention |
| Jul 2024 | Coder-V2 | 128K context, 338 programming languages |
| Dec 2024 | DeepSeek-V3 | 671B parameters, FP8 training, the $5.6M base |
| Jan 2025 | DeepSeek-R1 | Reasoning from pure RL; the App Store #1 moment |
| Aug 2025 | V3.1 | Hybrid thinking — one model, two modes |
| Sep 2025+ | V3.2 series | DSA sparse attention; Speciale reaches IMO/IOI gold level |
| Apr 2026 | V4 Flash and Pro | Native 1M context, CSA+HCA attention, MIT license |
| Aug 2026 | V4-Pro goes official | Agent upgrades, native Responses API, experimental vision |
Three threads run through the table. It started with code, not chat. It bet on sparse activation at V2 and never looked back. And in 2025 it swung hard at reasoning — R1 — which is the release the world noticed.
Why It's So Cheap: The Architecture
How does DeepSeek work? The core answer is one word: MoE, mixture of experts. Picture a hospital with 671 billion parameters of staff, where each patient — each token — lights up only the relevant departments. V3's 671B activates 37B at a time, under six percent. Model depth is huge; the compute bill is small. That's where the whole product line's pricing comes from.
Three supporting moves around the core:
- MLA (multi-head latent attention) compresses long-conversation memory into a low-dimensional representation — holding a long session stops costing a fortune.
- FP8 mixed precision trains with 8-bit floats where it's safe, saving memory and compute.
- CSA+HCA attention (V4) is the current frontier: fine-grained compression for new tokens, aggressive compression for old ones, and a fast indexer picking what to read. At a full 1M-token context, V4-Pro's inference FLOPs are 27% of V3.2's, with roughly a tenth the KV cache.
No magic — just engineering thrift, sustained for three years straight.
R1: The Pure-Reinforcement Experiment That Worked
In January 2025 DeepSeek ran what I still think is its most interesting experiment. Take a base model, train it with nothing but reinforcement learning and verifiable rewards — math answers are checkably right or wrong, code either runs or doesn't — and give it no human-written reasoning examples to imitate. The reasoning chains grew on their own. The paper later passed peer review and landed in Nature — the first major LLM to get there.
That RL stage cost about $294,000, on top of the $5.6M V3 base (2.788 million H800 GPU-hours). Set that against the $50M–$100M+ training estimates of the GPT-4 era and you understand the market's reaction — Nvidia dropped 17% in a single day when R1 landed. I watched the numbers roll in that afternoon; it was the largest one-day market-cap loss on record at the time.
R1 also left a finding the community ran with: its reasoning can be distilled into small models. Fine-tuning Llama and Qwen bases (1.5B to 70B, six models) on 800,000 R1 reasoning samples taught the small models to reason — and it worked better than running large-scale RL directly on them.
Benchmarks: Read Them with Dates
Every DeepSeek benchmark needs its date attached, because this line moves fast. The famous snapshot, from R1's launch in January 2025:
| Benchmark | DeepSeek-R1 | OpenAI o1-1217 |
|---|---|---|
| AIME 2024 (math) | 79.8 | 79.2 |
| MATH-500 | 97.3 | 96.4 |
| Codeforces (percentile) | 96.3 | 96.6 |
| SWE-bench Verified | 49.2 | 48.9 |
| GPQA Diamond | 71.5 | 75.7 |
| MMLU | 90.8 | 91.8 |
Roughly a tie with o1 at a fraction of the serving cost — that was the early-2025 story. The current generation: V3.2 sits in the GPT-5-High neighborhood, V3.2-Speciale reached 2025 IMO/IOI gold-medal level, and V4-Pro-Max scores 80.6% on SWE-bench Verified — the highest open-weight score, tied with Gemini 3.1 Pro. One honest caveat I keep repeating: V4's brightest numbers are self-reported and haven't yet been through independent reproduction. Trust the direction; treat the decimal points as provisional. The V4 deep-dive takes them apart.
What the Models Cost
Training cost is history; API price is what you pay. Since August 16, 2026, pricing splits into peak and off-peak hours — weekday peaks at full price, everything else half (the table shows off-peak; peak doubles it):
| Model | Input $/M (miss / hit) | Output $/M |
|---|---|---|
| deepseek-v4-flash | $0.22 / $0.007 | $0.66 |
| deepseek-v4-pro | $0.66 / $0.022 | $1.98 |
Peak hours are 9:00–12:00 and 14:00–18:00 Beijing time on weekdays; the other 17 hours of the day come at half price. Work evenings in China and you're on off-peak automatically. Treat the official pricing page as the source of truth — this line has changed prices twice in four months, so any number in an article (mine included) ships with an expiry date.
A migration pitfall I hit myself: the old deepseek-chat and deepseek-reasoner endpoints retired on July 24, 2026. That week my old scripts started erroring and I hadn't touched a line of code. If it happens to you, it's not you — rename the models to deepseek-v4-flash or deepseek-v4-pro and the migration is done; the API guide has the full steps.
Open Weights: What You Actually Get
Most of this line is MIT-licensed — download, run, fine-tune, ship commercially, no permission needed. Hugging Face shows five-million-plus downloads and 500+ community derivatives as of May 2025, and both numbers only grow.
What self-hosting actually takes:
- Full V3/V4: an 8× H200-class node. Pro's 865GB is multi-node territory.
- V4-Flash at 160GB: a seriously equipped workstation, or a rented box.
- Distilled series (1.5B–70B): consumer GPUs; some laptops cope.
- Tooling: Ollama for personal chat, vLLM or BentoML for production.
Open weights are exactly what they sound like — the right to download the model files and run them on your own machines. And openness buys real community work: DeepScaleR kept doing RL on the 1.5B distill and pushed AIME pass rate from 28.8% to 43.1%; a Berkeley team reproduced R1-Zero-style training for under $30.
How to Use DeepSeek
Four paths, in the order I'd tell a friend to try them:
- The web chat — free, no key, chat.deepseek.com. Expert Mode runs V4-Pro, Instant Mode runs V4-Flash. Start here.
- The API — base URL unchanged (
https://api.deepseek.com), OpenAI format or Anthropic format, your pick. Claude Code integration is two environment variables. - Agent tools — Claude Code, GitHub Copilot, and OpenCode can all use DeepSeek as the backend; Harness is the official agent product, in developer preview.
- Self-hosting — as above; the people who need it know who they are.
My rule after two years on it: Flash by default, Pro only when I can measure Flash falling short. The two sit one point apart on independent indexes and three times apart on price.
The Specialist Families
Beyond the main line, three specialist families (all open):
| Family | What it's for |
|---|---|
| DeepSeek-Prover-V2 | Formal theorem proving in Lean 4 |
| Janus / Janus-Pro | Unified image understanding + generation |
| DeepSeek-OCR / OCR-2 | Optical character recognition |
So "can DeepSeek draw?" — the main LLMs are text-only; drawing is Janus's department. Janus is the special model that does image understanding and generation in one, released separately from the main line.
Frequently Asked Questions
Is DeepSeek free? The web chat and the mobile app are free. The API is pay-as-you-go at the table above. The weights are MIT — free to run yourself too.
What kind of company is it? Any relation to OpenAI? Not a subsidiary of anyone — an independent research company, founded by Liang Wenfeng and funded by High-Flyer. All models trained from scratch in-house, not fine-tunes of someone else's.
Chinese only? No. Hangzhou-made and strongest in Chinese, but its English benchmarks sit in the same top tier — see the R1-versus-o1 table above.
V4-Pro or V4-Flash? Scale (1.6T vs 284B parameters), price (3×), and about one measured point of quality. Defaults beat spec sheets: most people should be on Flash. Details in the V4 guide.
Can I wire it into my editor or terminal? Yes — the API speaks both OpenAI and Anthropic formats, and Claude Code, Copilot, and OpenCode plug in directly. Writing good prompts for agents is its own craft; the prompt guide covers it.
Where does the name come from? "Deep seek" — deep learning, and exploring. The V is the architecture generation; the R marks the reasoning line that started with R1.
That's the whole picture as of late August 2026: a small Hangzhou lab that reached the top tier on architecture instead of budget, opens most of it under MIT, and prices the rest near cost. Start with the free web chat for a feel, move to the API when you want it in your own tools — the DeepSeek hub has the rest of the ecosystem.