DeepSeek landed in most people’s feeds in late January 2025, when its app shoved past ChatGPT on the US App Store and helped wipe roughly $589 billion off Nvidia’s market value in a single session. A Hangzhou startup that most Western developers had never heard of had just shipped something the big labs kept saying was brutally expensive: a frontier-class model with a reported training bill of $5.6 million.
That number deserves an asterisk, and I’ll get to it. But a year later, DeepSeek AI is more than a stock-market anecdote. It’s a family of open-weight models people genuinely run on their own hardware, an API that undercuts almost every competitor on price, and a fairly brutal lesson in how quickly the economics of this field can move.
What DeepSeek AI actually is
DeepSeek is a company spun out of High-Flyer, a Chinese quantitative hedge fund, and it’s led by co-founder Liang Wenfeng. The lab released DeepSeek-V2 in May 2024, V3 that December, and R1 in January 2025. All of them are mixture-of-experts architectures, which means only a slice of the network fires for any given token. V3 holds 671 billion parameters but activates around 37 billion per token. That’s the trick that let the team train on a cluster of 2,048 H800 GPUs rather than the tens of thousands the American labs were burning through.
R1 is the release that changed the conversation. It’s a reasoning model in the same family as OpenAI’s o1, and DeepSeek published the weights under an MIT license, which permits commercial use with almost no strings. If you want the fuller picture of where this class of model genuinely helps and where it still trips over itself, there’s a good breakdown of what reasoning AI does well and where it falls apart that’s worth reading alongside this.
The distilled versions are the real gift
Alongside R1, DeepSeek shipped six distilled models built on Qwen and Llama backbones, running from 1.5B to 70B parameters. The 7B and 8B versions will run on a laptop with enough RAM, and they hold up surprisingly well on math and logic problems. For anyone who wanted to study how reasoning training actually works without renting a data center, that was the practical doorway in.
Where the $5.6 million figure comes from, and where it doesn’t
The V3 technical paper lists about 2.79 million H800 GPU-hours for the final training run. Multiply that by going cloud rental rates and you land near $5.6 million. What the figure leaves out: the failed experiments, the salaries of a research team that reportedly pulled talent from ByteDance and Alibaba, and the years of infrastructure work that came before the run started. Analysis firm SemiAnalysis put the true all-in cost closer to $1.3 billion once hardware and staff are counted.
So the headline was misleading. The underlying point survives anyway. DeepSeek got a model into roughly the same performance band as GPT-4o for a training bill a mid-sized company could plausibly finance.
What DeepSeek is genuinely good at
- Code generation. The DeepSeek-Coder line was strong before R1 existed, and V3 kept that reputation. In side-by-side tests on Python refactors and SQL wrangling, it holds its own against models that cost twenty times more per token.
- Math and structured reasoning. R1 shows its chain of thought by default, which makes it easy to spot where a calculation went sideways instead of just getting a wrong answer with no explanation.
- Bulk inference on a budget. Cache-hit input pricing on the V3 API dropped to fractions of a cent per million tokens. If you’re classifying 400,000 support tickets or summarising a document archive, that difference is the whole business case.
- Fine-tuning. Open weights mean you can train on your own data without shipping it to anyone’s server. That matters for regulated industries in a way a cheaper API never will.
For teams building coding agents, the cheap token cost changes what’s feasible. There’s a decent overview of the current tools, models and agents in AI programming that puts DeepSeek’s pricing in context against the alternatives.
Running it yourself versus using the app
This is the fork in the road most people miss. The DeepSeek app and the deepseek.com API are one product with one set of rules. The weights on Hugging Face are a different thing entirely.
Self-hosting V3 in full precision needs something in the region of 600GB of VRAM, so realistically that’s a multi-GPU server or a hosted endpoint. Quantised versions and the distilled models are far more forgiving. A 32B distill quantised to 4-bit fits on a single 24GB consumer card, and it’s good enough for a lot of internal tooling. If you’re weighing that against other open models, this piece on what open source AI you can realistically run, own and ship is a useful reality check on hardware requirements.
The catch: privacy, censorship, and jurisdiction
DeepSeek’s privacy policy states that data collected through the app is stored on servers in China. That single line triggered a wave of government action. Italy’s data protection authority blocked the app outright in January 2025. Australia and South Korea restricted it on government devices, and several US federal agencies followed. Those aren’t abstract concerns; if you paste client contracts or patient notes into a chatbot, where the data lands is a compliance question, not a preference.
There’s a second issue that gets less airtime. R1’s alignment training produces noticeably cautious answers on politically sensitive topics, and researchers found the model sometimes reasons its way toward an answer and then declines to give it. Self-hosting the weights removes the app layer and the logging, but it doesn’t remove the behaviour baked into the model during training.
Neither problem is disqualifying. Both are things to know before you build a workflow on top of it.
How it compares to ChatGPT in practice
DeepSeek is not a drop-in ChatGPT replacement, and pretending otherwise sets you up for disappointment. It doesn’t have the plugin ecosystem, the image generation, the voice mode, or the polished agent tooling. What it has is raw text and code performance at a price that makes experimentation cheap.
My honest read after using both: for a one-off task where quality is everything, frontier closed models still win more often than not, especially on long multi-step instructions. For high-volume, well-specified work, the cost gap is hard to argue with. There’s a longer comparison of what ChatGPT and OpenAI actually deliver in 2025 if you want the other side of the ledger.
Getting real work out of DeepSeek
A few things I’d tell anyone starting now. Use R1 for problems with a checkable answer, like maths, unit conversions, or debugging a stack trace, and use V3 for anything chatty or creative where the visible reasoning just wastes tokens and money. Set a token budget on R1 calls, because it will happily think for 8,000 tokens on a question that needed 300.
If you’re on the app rather than the API, assume everything you type is stored. Treat it like a public forum, not a private notebook. For anything sensitive, pull down the weights and run the distilled model locally, even if the quality drops a notch.
Watch the pricing war, too. DeepSeek’s V3.2-Exp release in September 2025 introduced sparse attention and cut API prices by more than half overnight, and competitors followed within weeks. The model quality gets the headlines, but the price drops are what actually reshaped how teams budget for AI. A year ago, running a million tokens through a frontier model was a line item you justified to finance. Today it’s rounding error on a coffee budget, and DeepSeek is a big part of the reason why.