Open-weight models let anyone download a trained AI and run it on their own hardware, but what does that really mean?
In an AI model, what are the "weights"?
The program code that defines how the model runs
The amount of text the model can read at once
The full dataset the model was trained on
The learned numbers that set the model's behavior
Right! Weights are billions of learned numerical parameters; releasing them lets anyone run the trained model.
Weights are the learned numbers that set the model's behavior. Code, data and context window are separate things.
What is the common name for shrinking a model by storing its weights with fewer bits, such as 4-bit?
Fine-tuning
Quantization
Low-rank adaptation
Distillation scaling
Correct! Quantization stores each weight at lower precision, cutting memory use with a small quality loss.
It's quantization: storing weights at lower precision, like 4 or 8 bits instead of 16, to save memory.
A company releases a model's weights but not its training data or training code. How is it best described?
Fully open source, since anyone can download it
Free for any commercial use, by definition
Open-weight, but not fully open source
Open source, because weights count as source code
Yes! Open-weight means the trained parameters are shared; full open source would also include data and code to reproduce it.
It's open-weight but not fully open source. Without the data and training code, nobody can rebuild the model.
Why do many hospitals and law firms like self-hosted models?
Prompts are still logged by the model maker
They need constant internet to generate replies
Bigger models are always more accurate for them
Sensitive data can stay on their own machines
Exactly! When the model runs on your own hardware, confidential data never has to leave your network.
The main draw is that sensitive data can stay on their own machines, since nothing is sent to an outside provider.
Scratch the card to reveal a fact about open models
Hugging Face hosts well over a million models, most of them free to download and run yourself.
A 7-billion-parameter model is stored at 16 bits (2 bytes) per weight. Roughly how much memory do the weights need?
About 3.5 GB
About 7 GB
About 14 GB
About 56 GB
Right! 7 billion × 2 bytes ≈ 14 GB, before extra memory for the context cache.
It's about 14 GB: 7 billion weights × 2 bytes each. Precision matters as much as parameter count.
You quantize that same 7B model from 16-bit to 4-bit. About how big are the weights now?
About 28 GB
Correct! 4 bits is half a byte, so 7 billion × 0.5 bytes ≈ 3.5 GB, a quarter of the 16-bit size.
It's about 3.5 GB. Going from 16 to 4 bits cuts each weight to a quarter, so 14 GB becomes 3.5 GB.
A startup wants to sell an app built on an open-weight model. What should it check first?
Nothing, since open weights allow any commercial use
Only whether the model passed its safety tests
The model's license terms for commercial use
Whether the training data is included in the download
Yes! Licenses vary: some are permissive like Apache 2.0, others limit use by large companies or certain purposes.
The license comes first. Open-weight licenses differ a lot, and some restrict commercial use or user numbers.
Flip the card to reveal a fact about local AI
The llama.cpp project made it possible to run large language models on ordinary laptops and even phones.
Did you know?
Local models can run offline
Open weights always include training data
Quantization reduces memory use
LoRA retrains the whole model
GPUs speed up text generation
You want a model to answer in your company's style, with one GPU and little time. Which approach fits best?
Retrain the whole model from scratch on your text
Fine-tune all billions of weights at full precision
Train a LoRA adapter on examples of your style
Switch to the largest model you can download
Great choice! LoRA trains small add-on matrices while the base weights stay frozen, so it fits on modest hardware.
A LoRA adapter fits best: it trains a small set of extra weights instead of retraining or updating the whole model.
On a flight with no Wi-Fi, can a model already downloaded to your laptop still answer questions?
No, it must contact its maker's servers to think
Yes, because it runs fully on the laptop
No, it only stores the questions until you reconnect
Only if the laptop has no graphics chip at all
Yes! Once the weights are on your device, all computation happens locally, no internet needed.
It can: a local model runs entirely on the laptop, so it works offline once downloaded.
A 13B model runs slowly on a laptop, but quickly on a desktop with a big GPU. What mainly explains it?
The CPU and GPU are equally fast at this work
The desktop has a longer context window built in
The 13B model is smarter on the desktop
GPU memory bandwidth and parallel math speed
Right! Generating text means streaming all weights through memory for each token, which GPUs do far faster.
It's the GPU's memory bandwidth and parallel math. Each new token needs the weights read again, so fast memory wins.
Drag the slider to reveal a fact
Can a chatbot run on a tiny computer?
Hobbyists have run small language models on a Raspberry Pi, slowly but entirely offline.
A small 8B model beats an older 70B model on many tests. What does this best show?
Training data and methods matter, not only size
The tests must have measured context length only
Parameter count is all that decides quality
Smaller models always store more training data
Exactly! Better data, longer training and newer techniques can let a smaller model outperform a bigger, older one.
It shows training data and methods matter, not only size. Parameter count is just one factor in quality.
Which claim about open-weight models holds up best?
They are safe by default since anyone can inspect them
They let outsiders study and adapt models, but safeguards can be removed
They show their full training data to anyone who downloads them
They send usage data back to improve the original model
Right! Open weights help research and customization, but users can also fine-tune away safety behavior.
The strongest claim: they allow study and adaptation, yet safeguards can be removed. Openness alone doesn't guarantee safety.
A clinic wants a private chatbot that answers from its own policy files, on one 24 GB GPU. Which plan would work?
Run a 70B model at 16-bit and paste all files in
Pretrain a new model from scratch on the policy files
Use a hosted chatbot that keeps prompts on its servers
Run a 4-bit 8B model locally with document retrieval
Smart plan! A quantized 8B model fits easily in 24 GB, and retrieval feeds it relevant policy passages while data stays private.
A 4-bit 8B model with document retrieval works: it fits the GPU, keeps data local, and looks up the clinic's files when answering.
How did you like this quiz?
Perfect! Every answer right!
14 of 14 right
100%
Challenge your tech-savvy friends to take the quiz!
Without one, it says “A friend.”
Please choose another nickname.
Question 1: In an AI model, what are the "weights"?
Question 2: What is the common name for shrinking a model by storing its weights with fewer bits, such as 4-bit?
Question 3: A company releases a model's weights but not its training data or training code. How is it best described?
Question 4: Why do many hospitals and law firms like self-hosted models?
Question 5: A 7-billion-parameter model is stored at 16 bits (2 bytes) per weight. Roughly how much memory do the weights need?
Question 6: You quantize that same 7B model from 16-bit to 4-bit. About how big are the weights now?
Question 7: A startup wants to sell an app built on an open-weight model. What should it check first?
Lightning round 1
Question 8: You want a model to answer in your company's style, with one GPU and little time. Which approach fits best?
Question 9: On a flight with no Wi-Fi, can a model already downloaded to your laptop still answer questions?
Question 10: A 13B model runs slowly on a laptop, but quickly on a desktop with a big GPU. What mainly explains it?
Question 11: A small 8B model beats an older 70B model on many tests. What does this best show?
Question 12: Which claim about open-weight models holds up best?
Question 13: A clinic wants a private chatbot that answers from its own policy files, on one 24 GB GPU. Which plan would work?
Made with AI. This quiz was made with AI, which can make mistakes.