DeepSeek Has Been Free for a Year and a Half. This Month It Got Hands.

DeepSeek published its first free model in January 2025. Not a trial, not a discount. The real thing, released under a licence called MIT that lets anyone download it, use it commercially, and owe nothing. They have kept doing it ever since. The model they released this summer is free on the same terms.

So the free part is not news, and anyone telling you it is has not been paying attention for eighteen months.

What changed on the thirteenth of August is something else. DeepSeek released the piece that lets the model work inside your computer instead of behind a chat window. That is the part worth understanding, because it moves DeepSeek into the same category as the tools I use every day.

Chatting with a model, versus giving it hands

Think about the difference between a consultant you can telephone and a consultant who comes to the office.

On the phone, they are just as clever. You describe your problem, they talk you through it, you write down what they said, and you go and do it yourself. That is a chatbot. That is DeepSeek’s website, ChatGPT, and every AI most people have used.

The consultant who comes in is a different arrangement. They open the filing cabinet, read the actual documents, change them, and hand them back done. Same person, same knowledge, completely different value to you, because they touched the work instead of describing it.

The second kind is called an agent. It runs on your machine, reads your real files, and carries out the task rather than explaining it.

The piece everybody was missing

Here is what was strange about the last eighteen months. The free models were excellent, and almost nobody outside of engineering could use them for real work.

The reason is that a model on its own cannot do anything. It answers one question at a time and then forgets. It cannot open a file, run a command, or keep track of what it did an hour ago. All of that comes from software built around the model: the part that hands it your files, keeps the notes, and decides when it is finished.

If the model is an engine, that surrounding software is the rest of the car. The steering, the pedals, the fuel line.

Companies gave away engines and kept the cars. Anthropic built Claude Code, OpenAI built Codex, and those tools are what turned a clever model into something that does your work. The models were open. The thing that made them useful was not.

On the thirteenth of August, DeepSeek published theirs. It is called dsh, it runs in your terminal, and it belongs to the same family as Claude Code and Codex. You type in plain English, and it reads and edits your actual files.

So is it free for your own use

This is the question I went in with, and the answer has two halves.

The tool is free. Genuinely, no conditions, no account, the same generous licence as the model. You install it and it is yours, including for client work.

The thinking is not. dsh is the steering wheel, and there is no engine inside it. On first run it asks for a key to a paid account, because your request has to be sent somewhere that can actually think. That somewhere is DeepSeek’s own computers, and they bill you for the trip. It is inexpensive as these things go. It is still a meter.

The licence costs nothing and the hardware costs plenty, and those two facts get blurred together every time someone calls a model free.

There is a third option, and it is the only fully free one. Download the model and run it on your own machine, so nothing leaves your house and nothing is billed. That requires a dedicated graphics card, the component that makes video games look good, in a considerably larger size than most people own. My desktop does not have one, so I cannot do it.

Why owning the file still matters

On the twenty-fourth of July this year, DeepSeek switched off two of its older API names. Anyone whose work depended on calling deepseek-reasoner had that door close on a scheduled date, announced three months in advance.

The model behind it, R1, is still sitting on Hugging Face, free, under the same licence it launched with in January 2025. Nineteen months later you can still download it.

Same company, same model, two different relationships. The rented version ended on a date somebody else chose. The owned version did not, and cannot, because it is a file on a drive and there is no mechanism by which anyone can reach in and remove it.

That is the actual argument for open models. It has never really been about the price.

Where it stops

If you want to try it this week, the honest order is this. Use the website first, because it costs nothing and takes two minutes. If it earns its place, install the terminal tool and attach a paid account, and accept that you are renting the thinking.

And do not put it under anything that matters yet. The version number is 0.1.0-rc.7, which means it has not reached its first proper release, and the project carries a written warning that updates will break things.

The free model has been available since the January before last. What arrived this month is the steering wheel, and it is still a prototype.

Forward → Upward ↑ Onward ↗︎
Mstimaj

Sources and Further Reading

• DeepSeek API release history and deprecation notices. Records the V4-Pro GA release on 2026-08-13 and the discontinuation of the deepseek-chat and deepseek-reasoner model names on 2026-07-24.
• DeepSeek-R1 release announcement, 20 January 2025. “Code and models are released under the MIT License.”
• deepseek-harness repository, GitHub. The terminal agent, MIT licensed, published as a developer preview.
• DeepSeek-V4-Pro model card, Hugging Face. “This repository and the model weights are licensed under the MIT License.”
• @deepseek-ai/dsh package listing, npm. Version 0.1.0-rc.7, MIT.