Abed ManzaljiAI, Step by Step

What's an open-weight model?

An AI model whose brain you can download and run on your own computers.

3 min to watch2 min to read

Based on real events: DeepSeek released R1 on Jan 20, 2025,1 Nvidia lost $589 billion in value on Jan 27,2 and Italy blocked DeepSeek's app on Jan 30.3 Where the app stores data comes from its privacy policy.4 The team, the files and the screens are made up.

Read the walkthrough as text

In DeepSeek's app

  1. January 20, 2025. A company called DeepSeek shares its new AI model, R1, for anyone to download.
  2. What you download are its weights: the 671 billion numbers it learned in training.
  3. This is an open-weight model: anyone can download it and run it on their own computers. Its MIT license allows work use.
  4. A week later, on January 27, Nvidia loses about $589 billion in value in one day. Investors fear AI may need fewer expensive chips.
  5. Back at work, your team wants R1 to summarize private customer files. The quickest way skips the download: DeepSeek's own app.
  6. You paste the files into the chat.
  7. To read them, the app sends your files to DeepSeek's computers.
  8. DeepSeek's privacy policy says what you send the app, files included, is stored on servers in China.
  9. January 30, 2025: Italy's data protection authority blocks DeepSeek's app to protect people's data.

Same model, wrong place to run it. In DeepSeek's app, what you send is stored on servers in China, says its privacy policy. Italy's data protection authority blocked the app on Jan 30, 2025.

On your own servers

  1. Same model, other path. This time your team downloads the weights to its own server.
  2. First, the license. R1's is MIT: you may use it at work, change it and build on it.
  3. The catch is the hardware. Every weight must sit in GPU memory, on the chips that run AI.
  4. R1 also comes in smaller versions. The 32B one needs about 2 GPUs.
  5. On your GPU server, one command loads the weights onto those 2 GPUs and starts answering requests.
  6. Your team pastes the same files into its own chat app.
  7. R1 answers from your own server. The files never leave your company's network.

That's what open weights give you. You choose where the model runs and who sees the data. Run from downloaded weights, R1 sends nothing back to DeepSeek. In return, the hardware is your job.

Keep going:

Why it matters, what it costs, and what to ask your team.

The real case

An open model shook the stock market.

An AI model is billions of numbers learned in training. Those numbers are its weights. Open weights means you can download them and run the model yourself.

On January 20, 2025, DeepSeek did exactly that with R1: 671 billion weights, MIT license, results it said matched OpenAI's o1.1,6 A week later, Nvidia lost more market value in one day than any company ever had.2,5

$589Bof Nvidia's market value gone on Jan 27, 2025
17%drop in Nvidia's share price that day
0days after R1's weights went public

Investors worried that if top AI could be built this cheaply, companies might need fewer expensive chips.2

Technically, I'm just very big files full of numbers.

Rent or own?

Which fits your team?

With an API, you rent: you send text to someone else's model and pay per use. With open weights, you own a copy: you pick the hardware, the version and who sees the data.

What's true for your team? Tap all that apply.

Rent an APIRun open weights
← Rent an APIRun open weights →

Longer bar = better for you. Ratings are a rough guide, not measurements.

Free to download isn't free to run. Someone buys or rents the GPUs, and someone keeps them running.

So free to download doesn't mean free to run.

Ask your team

Three questions before you choose.

The model is the easy part. Where it runs is the real decision.

  1. What data will the model see, and where may that data go?Often this one answer decides it: a rented API, or your own servers.
  2. What does it cost at our real volume, all in?An API bills per use. Your own GPUs cost the same busy or idle, plus the people who run them.
  3. What does the license allow?"Open" ranges from MIT, which allows almost anything, to licenses with user limits and naming rules.10

Under the hood, step by step: run an open model on your own server, so private files never leave.

Build along

Run R1 on your own server, one step at a time

We start with the quickest version and give it the same customer files every time. Press Run it after each step and watch where the files go.

Before you ship

A short checklist.

Open weights come with fine print, and with chores an API used to do for you.

  1. Read the license, then the base model's licenseR1 itself is MIT. Its small versions are built on Qwen or Llama, and those licenses still apply.6
  2. Check the limitsLlama 3.1, for example, needs a separate license from Meta above 700 million monthly users, and a "Built with Llama" notice.10
  3. Test the exact version you'll shipRun your evals on the same size and bit-width. A shrunk model can slip on your tasks.
  4. Lock the doorRequire an API key, keep the server off the public internet, and log every request.
  5. Pin the versionKeep a copy of the exact weights, so the model only changes when you decide.

References

  1. "DeepSeek-R1 Release" · DeepSeek, Jan 20, 2025. Open weights under the MIT license; claimed on par with OpenAI o1 on math, code and reasoning.
  2. "Nvidia stock begins recovery after DeepSeek AI frenzy prompted near $600 billion loss" · Laura Bratton, Yahoo Finance, Jan 28, 2025. A 17% fall and a $589 billion market-cap loss on Jan 27, the largest single-day loss in stock market history; why investors worried.
  3. "Italy blocks access to the Chinese AI application DeepSeek to protect users' data" · The Associated Press, Jan 30, 2025. Italy's data protection authority (Garante) blocked the app.
  4. "Is DeepSeek really sending data to China? Let's decode" · Shubham Sharma, VentureBeat, Jan 27, 2025. The app's privacy policy: prompts, uploaded files and chat history are stored on servers in China; a self-hosted R1 does not send data there.
  5. "Nvidia Sheds Nearly $600 Billion in Historic Single-Day Drop" · Felix Richter, Statista, Jan 28, 2025. $589 billion on Jan 27, 2025.
  6. deepseek-ai/DeepSeek-R1 model card · Hugging Face. 671B total weights, stored in 8 bits (FP8); MIT license, commercial use allowed; distilled sizes from 1.5B to 70B built on Qwen2.5 (originally Apache 2.0) and Llama bases, the 32B one on Qwen2.5-32B; the vLLM command and the 0.6 temperature tip.
  7. "Quantization" · Hugging Face Optimum docs. Lower-precision numbers need less memory, often with some loss in accuracy.
  8. NVIDIA H100 Tensor Core GPU · NVIDIA. 80 GB of GPU memory (H100 SXM).
  9. "OpenAI-Compatible Server" · vLLM docs. vllm serve, --api-key, port 8000, and the OpenAI client example.
  10. Llama 3.1 Community License Agreement · Meta, Jul 23, 2024. The 700 million monthly-user threshold and the "Built with Llama" notice.
  11. deepseek-ai/DeepSeek-R1-Distill-Qwen-32B model card · Hugging Face. R1's 32B version, stored in 16 bits (BF16).

The team, the files, the chats and the screens are made up; the dates, the $589 billion, the Italy block and what the privacy policy says are real. The GPU numbers are rough math (weights at the bits they're published in, plus 30% room to work, 80 GB per GPU), not measurements. The rent-or-own ratings are a rough guide.

Made by Abed Manzalji, AI Engineer in Toronto.More about meContact