All Things AI · Season 2 · The Models · Episode 4
DeepSeek is an open model, published under the MIT license, that costs a fraction of the other paid services to run and is too large to run on a laptop. The practical version is hosted. We run it through NVIDIA’s API on free trial credits, which keeps the file on NVIDIA’s servers rather than DeepSeek’s and the monthly bill at zero.
What it is
A family of open models from a company in China, published with the weights anyone can download and a license, MIT, that lets you use them for anything. The current generation is DeepSeek-V4, and it comes in two sizes. Flash is 284 billion parameters with 13 billion active on any given word. Pro is 1.6 trillion parameters with 49 billion active. Both read up to a million tokens at once, a limit few working documents come anywhere near.
The thing to understand first is the word open. Hermes and GLM are open models you download and run on your own machine. DeepSeek is open in the same sense, but the models are so large that running one yourself means a rack of server hardware, not a laptop. So for an agent, open means something different here: several companies can host the same model and sell access to it, and you get to choose which one, at what price, in which country.
Why the price is the point
Through DeepSeek’s own service, the Flash model costs about twenty-two cents per million tokens in and sixty-six cents per million out during off-peak hours, roughly double that at peak. Pro is three times Flash. At those rates a long follow-up draft costs a fraction of a cent.
That changes behaviour more than it changes budgets. When a draft costs nothing measurable, you stop deciding which twenty of the hundred follow-ups deserve a proper attempt and you run all hundred, then spend your time correcting rather than producing. The episode on Grok makes the same argument about speed. This is the same argument about price, and among the models with a per-token bill in this season, DeepSeek is the cheapest by a wide margin.
The two ways to run it, and why we use the second
Because the weights are open, there is more than one door.
DeepSeek’s own service
The company sells access directly, at the prices above, through the same style of connection every other model uses, so any tool built for one of the paid services can be pointed at it by changing one address. It is DeepSeek’s own infrastructure serving the same V4-Flash and V4-Pro models described on its own model card, at the prices set on its own pricing page.
It is also the door where the file goes to China. DeepSeek’s own privacy policy, updated in February 2026, says the data it collects, including prompts and uploaded files, is stored in the People’s Republic of China, and that it may be used to train its models unless you opt out. That is not a rumour. It is their document. For a client’s material, that settles the question.
NVIDIA’s hosted version
NVIDIA runs the same open weights on its own hardware and offers them through its developer platform as a trial service, governed by the NVIDIA API Trial Terms of Service: you make an account, get a key, and receive free credits. Both the Flash and the Pro models are on it, running on NVIDIA’s servers rather than DeepSeek’s.
This is how we run it right now. The tokens are free, the connection works the same way as every other model, and the material goes to NVIDIA’s service, not DeepSeek’s. NVIDIA does not publish a specific rate limit for the trial; its terms describe it only as a trial service under the NVIDIA API Trial Terms of Service. For the way an agent actually uses this, a batch of drafts in the morning and a few dozen through the day, it has not been a constraint.
What to use it for
The profile is volume, length, and anything where the cost of running the job was the reason it did not get done.
- Follow-up drafts in bulk, from your own notes, for you to correct rather than compose.
- Long documents. A million tokens of context means the whole inspection report, the whole set of disclosures, or a year of your own posts fits in one request.
- Second drafts. Running your material through a second, differently trained model is a cheap way to get a genuinely different attempt rather than the same one reworded.
- The repetitive middle of any pipeline you build, where a model is doing the same small job hundreds of times and the price per call is what decides whether the pipeline is worth having.
- Anything where you want a thinking mode you can switch off. It reasons at length when asked to and answers quickly when not, and the quick mode is the right one for routine drafting.
Where it is weaker
Three honest limits.
You cannot run it on your own computer, and pretending otherwise wastes an afternoon. The smaller open models cover that job. This one is for a hosted service, and the question is only which host.
The Flash model is weaker at recalling facts than the Pro model, and the company’s own model card shows it plainly. For a draft from your own material that does not matter, because the facts are in the material. For a question that relies on what the model knows, use Pro or check the answer.
On its own, through the API, it does not look anything up. It knows what it was trained on. Anything that depends on what is true this week comes from your files or from a tool that is genuinely online.
Where this goes wrong
Three failures, and the first one is the serious one.
Sending a client’s file through DeepSeek’s own service because it is cheap. The privacy policy is unambiguous about where that goes. Use a host whose privacy terms you have read, such as NVIDIA’s, for anything with a client in it, or use one of the models that never leaves your laptop.
Leaving thinking mode on for everything. It is on by default, it costs output tokens, and for a two-paragraph follow-up it adds time and money for nothing. Switch it off for routine work and on for the hard problem.
Building a daily pipeline on a trial key and then being surprised when it slows down. Free credits are for finding out whether this fits how you work. Once it does, a paid host at these prices is still close to nothing.
What to try in the first week
Two exercises, both chosen to make the price and the size concrete.
Take the longest document you have on hand, something you have been meaning to read, and ask for every date, deadline, and dollar figure with the page it came from. Then look at what that request cost. The number is the demonstration, and it is the moment most agents realise they have been rationing something that did not need rationing.
Then take thirty sets of notes from thirty conversations and ask for thirty follow-ups in your usual shape. Do not read them for quality yet. Read them for how many you would send with one correction. That ratio is what this model is for, and it is the ratio to compare against whatever you pay for today.
How it sits next to the others
Against Hermes and GLM it is the open model you cannot run yourself, in exchange for being considerably larger and more capable than anything that fits on a laptop.
Against the paid services it is the price. The capability is in the same conversation, and the cost is a different conversation entirely.
The natural arrangement is a local model for anything with a client in it, DeepSeek through a host you have chosen for volume, and one paid service kept for the small number of jobs where you want a specific product rather than a model. Which one to pick, the last episode of this season, walks through that split.
Why an open model this size matters
It is worth saying what is unusual here, because it is easy to miss.
The largest models used to belong to the companies that trained them, and you could only reach them through that company’s door at that company’s price. An open model at this scale means the same capability is available from several hosts, at several prices, under terms you can read, and none of them can raise the price or change the model underneath you without you being able to walk to the next one.
For an agent that is the difference between renting a capability and choosing a supplier for one. The episodes in Season 3 on building your own tools depend on exactly this: a model you can call from anywhere, that costs almost nothing per call, that you are not locked to.
What it costs
Through NVIDIA’s trial, nothing, under its published trial terms. Through DeepSeek’s own service, about twenty-two to forty-four cents per million tokens in and sixty-six cents to a dollar thirty-two out for Flash, three times that for Pro, and the company halves its rates outside its peak hours. Other hosts sell the same open weights at their own prices. Whichever door you choose, the cost question is settled and the only real question is where the file goes.
Questions agents actually type
Is DeepSeek free?
The model is free to download under the MIT license, but it is too large to run on a laptop. Through NVIDIA’s developer platform you can use it on free trial credits under NVIDIA’s published trial terms. Through DeepSeek’s own service it is paid, at about twenty-two to forty-four cents per million tokens in for the Flash model, which is a fraction of what other paid services charge.
Is DeepSeek safe for real estate client files?
Not through DeepSeek’s own app or service. Its privacy policy says data, including prompts and uploaded files, is stored in the People’s Republic of China and may be used for training unless you opt out. Use the same open model through a host whose privacy terms you have read, such as NVIDIA’s, or use a model that runs entirely on your own computer for anything with a client in it.
What is the difference between DeepSeek V4 Flash and V4 Pro?
Flash is 284 billion parameters with 13 billion active; Pro is 1.6 trillion with 49 billion active. Both read a million tokens of context. Pro is stronger at recalling facts and at the hardest reasoning and costs about three times as much. For drafting from your own material, Flash is usually enough.
Can I run DeepSeek on my own computer?
Not realistically. The current models need server-class hardware. If running locally is the requirement, the smaller open models covered in this season, Hermes and GLM, are the ones that fit on a laptop. DeepSeek is the open model you reach through a hosted service.
Season 1 is Claude Code for Real Estate Agents, filmed and on its own hub. All three seasons: AI for real estate agents.