DeepSeek is a Hangzhou lab with a free chat site, an API meter, and open-weight files on Hugging Face. Those are three different things to install, and mixing them up is expensive. Do not clone hundreds of gigabytes because a Slack message called the lab “free and open.”
Say a teammate in Slack tells you DeepSeek is “free and open,” and you go clone it late one night. The actual job is a six-line shipping FAQ: carrier cutoff 3 pm Pacific Time (PT), a signature required on orders over $250, no weekend pickups. Hugging Face shows deepseek-ai/DeepSeek-V4-Pro. The Files tab lists that checkpoint at roughly 865 gigabytes (GB). Your laptop has 412 GB free. Your computer throws a disk-full warning partway through the download, with over a third of the file still missing. The FAQ still says “TBD.” You spent half the night pulling a massive folder of model weights when all you needed was the website.
If you only needed a thank-you email, start at the product chooser and skip the Files tab entirely. If you needed the vocabulary for what “open” means with AI weights, read what open means for AI weights and licenses. If you needed hosted versus a file on your own disk, read hosted models versus downloading to your machine. This page covers DeepSeek’s own two paths in one place. More background lives on Learn.
Two paths, one Hangzhou lab
DeepSeek (深度求索) is a company founded in 2023 in Hangzhou, and it offers you two different ways to use its models. The first is the product: a website, phone apps, and an API that all run models on DeepSeek’s own servers. The second is the files: checkpoints published under the Hugging Face account deepseek-ai, which you can download if the license on that specific repository allows it. It is the same lab either way, but the disk space, the privacy story, and the hardware bill are completely different.
Some workplaces restrict DeepSeek for region or data-policy reasons, and that caution is real. It gets its own detailed post later in this series, not a geopolitics essay here. If your employer has not said yes, do not paste customer text into any of the three options below. Use another vendor from the product chooser instead, or wait for a clear answer.
People collapse three different things into the single word “DeepSeek”:
| Way in | Where it lives | What you run | Where your prompt goes |
|---|---|---|---|
| Chat app | chat.deepseek.com and the mobile apps | A browser or app. Instant and Expert modes on the V4 line, per the docs we checked in August 2026 | DeepSeek’s servers |
| API | api-docs.deepseek.com, keys issued on the platform site | Your script or agent, using model ids such as deepseek-v4-flash or deepseek-v4-pro | DeepSeek’s servers, billed in tokens rather than a chat seat |
| Open weights | huggingface.co/deepseek-ai | A checkpoint folder. V4-Pro’s Files tab held the 865 GB class when we checked in August 2026 | Your own machines if you stay local; a host if you then upload the files there instead |
Rule of thumb: If you never opened the Hugging Face Files tab, you are using the product. If you did open it, you are shopping for hardware.
The chat product is just a website
Open chat.deepseek.com, sign in, and type. That is the product you actually needed for the shipping FAQ. DeepSeek’s V4 preview post, from April 24, 2026, told people to try the models there in Expert Mode and Instant Mode. Mode names can change over time, but the site itself stays the door you walk through. There are also iOS and Android apps. Plans, usage caps, and whether the free tier stays free get covered in the next post in this series, so do not memorize a single price you saw in a tweet.
The chat product is hosted, which means your prompt leaves your laptop the moment you hit send. Retention, training use, and which country the servers sit in are DeepSeek’s policies, not something a local folder controls. Even if the FAQ had contained customer addresses, you would still need a simple paste test: would you be comfortable emailing this text to that vendor? A giant download you never actually load is not a privacy control, no matter how big the file is.
Use chat when the job is writing, answering questions, getting a first feel for how the model talks, or drafting something you plan to edit yourself. Do not use chat as proof that you “run DeepSeek.” What you are actually running is a browser tab.
The API is still their GPUs
The API is the same hosted family of models, unlocked with a key instead of a login screen. Official docs live at api-docs.deepseek.com. Checked in August 2026, the first-call page lists https://api.deepseek.com as the OpenAI-compatible base_url, plus an Anthropic-compatible path, and model ids deepseek-v4-flash, deepseek-v4-pro, and an experimental vision id. The docs also note that those flash and pro ids currently route to specific dated drops (Flash-0731, Pro-0813), so re-read that table the week you actually ship. Do not copy a snippet from 2025 that still says deepseek-chat.
DeepSeek’s V4 preview post said the older ids deepseek-chat and deepseek-reasoner would be fully retired after July 24, 2026, at 15:59 UTC. If a tutorial you’re reading still uses those exact strings, it is out of date, so swap the model field for a current one. Keep your key in an environment variable rather than in code. A chat-app login does not pay your token bill either; a chat seat and console credits are billed by different systems entirely.
Use the API when you are wiring up an app, running a batch rewrite job, or building an agent that already speaks OpenAI- or Anthropic-shaped requests. Use chat when you are a person with a plain FAQ to write. Mixing the two is how someone pastes an API key into a random plugin and then blames “DeepSeek” for a surprise usage invoice.
Open weights are files, not a chat box
The Hugging Face account deepseek-ai is the public shelf for DeepSeek’s released model checkpoints. As of our August 2026 check, the V4 collection includes Flash and Pro, in base and instruct versions, plus later dated cards such as DeepSeek-V4-Flash-0731 and DeepSeek-V4-Pro-0813. The preview Pro card is still the one whose Files tab shows 865 GB, and dated drops can run larger. Open the Files tab yourself and write the number down. That number tells you about disk space, not about how smart the model is.
V4 is what is called a Mixture of Experts (MoE) family, which means the network has many smaller expert blocks inside it, and only some of them switch on for any given word it processes. DeepSeek’s own V4 writeup lists V4-Pro at about 1.6 trillion total parameters, with roughly 49 billion of those active per word. V4-Flash runs about 284 billion total, with about 13 billion active. The active number is the actual work per word; the total number is why the download folder is so huge. A laptop that can comfortably chat with a small 8B GGUF file will not hold V4-Pro. Community-made compressed versions exist, but they are still a real hardware project, not a quick download.
The V4 model cards we checked in August 2026 mark the weights under a permissive open-source license called MIT (MIT). Read the actual LICENSE file on whichever repository you clone, because licenses on older DeepSeek releases were not always the same, and cards can change over time. “Open” here means open weights plus that specific license text. It does not mean the chat app’s own source code, the training data behind it, or a promise that your prompt stays on your machine. The linked guide above covers that vocabulary in more depth; this page only needs one point: the files are optional, and 865 GB is a hardware project, not a quick download.
You can also send those same downloaded files to a paid host, such as Together, Fireworks, or a graphics processing unit (GPU) you rent by the hour in the cloud. At that point you are back to a hosted setup, just running someone else’s servers instead of DeepSeek’s. Cloning the files from Hugging Face does not by itself make the prompt local. What actually makes it local is running the weights on hardware you control, with cloud syncing turned off.
What to call the V4 models
Treat every model name below as something that can shift. DeepSeek announced a V4 preview on April 24, 2026, covering DeepSeek-V4-Pro and DeepSeek-V4-Flash, both offering a 1 million token context window on official services. Later, they posted dated official drops (Flash-0731, Pro-0813). The API keeps the short ids and quietly points them at whichever drop is current. Chat interface labels, such as Instant, Expert, or a thinking mode toggle, are just a skin sitting on top of that same family. When in doubt, quote the exact id from the docs or the model card you opened today.
Here is what you can say without running your own bake-off: both V4 doors are Mixture of Experts models, both advertise a 1 million token context window on DeepSeek’s own stack, Flash is the smaller and cheaper hosted id, and Pro is the larger one. DeepSeek publishes benchmark tables on the model cards and in the April announcement, and this article will not reprint them, because scores shift, prompts differ, and a six-line shipping FAQ will never notice a two-point gap on a public leaderboard. If you genuinely need a comparison, run your own five prompts on both models and keep the outputs.
Older names you will still run into in blog posts include V3, V3.2, R1, various coder-named checkpoints, plus the ids deepseek-chat and deepseek-reasoner. Treat those as history unless a model card you opened today still lists one of them as current. It helps to keep a sticky note nearby with three words on it: chat site, API id, or Files tab.
Worked example: 865 GB, six lines, one Slack myth
The draft shipping FAQ, before anyone touched it:
# Door pick (DeepSeek, as of writing August 2026)
# Job: rewrite the six-line shipping FAQ (cutoff, signature, weekends).
# May this text leave the building: yes (public policy page)
# Work policy on DeepSeek (Hangzhou lab): allowed for public copy # DS6 if unclear
# Hardware we have: 16 GB laptop, 412 GB free disk
#
# Files tab on deepseek-ai/DeepSeek-V4-Pro: 865 GB class
# Decision: do NOT clone. Open chat.deepseek.com.
# If this were an app: API id deepseek-v4-flash, key in env.
# If this were self-host: stop; 865 GB > 412 GB free, and RAM would fail first.It is easy to skip that door-pick step and match the word “open” straight to “clone,” which is exactly what happened here. Free disk was 412 GB; the Files tab said 865 GB, so the clone failed partway through. Even if it had finished, a 16 GB laptop cannot hold a checkpoint from a 1.6-trillion-parameter MoE model, so the download was doomed either way. Chat would have drafted the six public lines in one sitting instead. You still would have had to type in the real 3 pm PT cutoff yourself, because no model anywhere knew that warehouse rule until you pasted it in.
| Claim in Slack | What you might try | What would have matched the job |
|---|---|---|
| “DeepSeek is free and open” | Cloned the V4-Pro weights | Free chat for a public FAQ, or read the LICENSE file if you truly needed the files |
| “Install DeepSeek” | Filled the laptop’s disk toward 865 GB | Opened chat.deepseek.com, or installed an SDK for the API with pip |
| “We run it ourselves” | A partial folder and a disk-full warning | Hardware you did not actually have; call it hosted if the prompt leaves your machine |
| Six-line rewrite | Still marked TBD hours later | A paste into chat, then a quick human edit of the real cutoff time |
The fix, in order: cancel the clone. Delete the partial shards if disk space is tight. Open the chat site on your work account. Paste in the six lines plus the real cutoff time. Edit the draft by hand. If the team later wants DeepSeek wired into a tool, get an API key and set deepseek-v4-flash from the current docs. If someone still wants local weights on their own hardware, they need a machine that can actually hold the Files-tab number, plus the location guarantees covered in the hosted-versus-download guide above. The FAQ itself does not need to wait for that purchase.
Do not clone 865 GB for a six-line FAQ
- Cloning V4-Pro because a Slack thread said “open,” when the actual job was one paragraph in a browser.
- Calling an open chat tab “self-hosted DeepSeek,” when nothing is running on your own hardware.
- Shipping code that still calls
deepseek-chatafter the July 2026 retirement date. - Assuming a chat login pays for API tokens, or assuming the reverse.
- Skipping your company’s work policy because the model card says MIT. MIT covers the files. The chat product is still a vendor based in Hangzhou, and that caution gets its own post later in this series.
- Reprinting a benchmark screenshot instead of just running the six lines you actually need through the model yourself.
Write the Files-tab number, then use the site
Write yourself a quick checklist for one real job you actually have. Fill in the Files-tab number from the card you would have cloned, even if you never clone it. Then do the job on chat.deepseek.com if your workplace allows it, or on the product chooser’s usual pick if it does not. The next post in this series covers chat app plans without the hype, so do not sign up for a paid plan until you have used the free site for a week. If what you actually wanted was hosted versus local in general, go back to the hosted-versus-download guide.
Chat, API, files
- DeepSeek is a Hangzhou lab that offers a chat product, an API, and open-weight files, and those are three separate things.
- Chat lives at chat.deepseek.com. The current API ids are deepseek-v4-flash and deepseek-v4-pro. Weights live on Hugging Face under deepseek-ai.
- V4-Pro’s Files tab is an 865 GB-class folder, and a six-line FAQ does not need any part of it.
- Work and region caution is real and gets its own post later in this series; always confirm live model names against DeepSeek’s own docs first.
Series notes
This is Part 1 of Learn DeepSeek. Next: chat app and plans without hype.
Sources
Vendor pages and docs checked in August 2026:
- DeepSeek Chat (the hosted chat product)
- DeepSeek API docs (first call, model ids, thinking mode)
- DeepSeek: V4 Preview Release (April 24, 2026: V4-Pro, V4-Flash, 1M context, legacy id retirement)
- Hugging Face: deepseek-ai (account, V4 collection, Files tab sizes)
- Hugging Face: DeepSeek-V4-Pro (MoE totals, 1M context, MIT license, 865 GB Files tab)
- Hugging Face: DeepSeek-V4-Flash (smaller V4 line)
- DeepSeek homepage
- AMS: Open-source AI explained and hosted versus download guide
- AMS: Learn and Which AI product should I use?
Keep going
Same lessons in your feed
Short diagrams, hooks, and weekly tutorials on Substack, Instagram, X, and Facebook.
