Quiz 12 min
Quiz: Open models
Six questions on the lesson Open models, each answer with its reason. The questions come from the pool of the module and are in the format of its exam.
Topics: Language models
Part of the lesson Open models
You turn off wi-fi, and the model in LM Studio still answers. What follows from this?
The whole model is stored on the laptop as a file and is computed by its processor or graphics card.
Correct: everything the model can do lies in the file with its parameters; nothing else is needed to answer.
LM Studio caches earlier answers from the internet and replays them while offline.
The model generates every answer anew, token by token; it also answers questions offline that were never asked before.
The model answers from a small database of texts that was installed with LM Studio.
A language model holds no texts to look up, but numbers from which it generates new text.
The model sends the questions once the connection is back and shows provisional answers until then.
A local model sends nothing anywhere; the answer it gives offline is already the final one.
A model has 8 billion parameters, each stored with 4 bits. About how big is the file?
4 GB
Correct: 8 billion times 4 bits is 32 billion bits, which is 4 billion bytes.
8 GB
That would be 8 bits, one byte per parameter. With 4 bits each parameter needs only half a byte.
32 GB
This confuses bits and bytes: 32 billion bits are 4 billion bytes, not 32.
16 GB
That corresponds to 16 bits per parameter, the unrounded storage; quantized to 4 bits it is a quarter of that.
On the same laptop, Gemma 3 4B answers noticeably more slowly than Gemma 3 1B. Why?
Every token passes through all parameters, and the 4B model has four times as many.
Correct: more numbers per token mean more computation per token and therefore fewer tokens per second.
The larger model searches a larger database of texts for every answer.
A model does not search stored texts; it computes every token from its parameters.
The larger model writes longer answers because it knows more.
Longer answers take longer in total but do not change the tokens per second, and those are lower for the larger model.
The larger model checks every answer on the internet before showing it.
Both models run offline; neither checks anything on the internet.
Which statements about open models are correct?
Their weights can be downloaded and run on your own hardware.
Correct: this is exactly what makes a model open.
Their training data is always published along with the weights.
Open weights are not open source; the training data usually stays secret, and only a few models such as Olmo publish it.
They can be adapted, for example fine-tuned on your own data.
Correct: whoever has the weights can change them further.
They may come with licences that restrict some uses.
Correct: open does not mean without conditions; some licences limit commercial use, for example.
They cannot be used inside an agent.
A harness can talk to any model, including a local open one.
In which situation is a local open model the most sensible choice?
A veterinary practice wants to summarise treatment notes that must not leave the practice.
Correct: when the data must not leave the building, a local model is often the only way.
A student wants the best possible answer to a difficult research question on a slow laptop.
A model that runs on a slow laptop is small; for the best possible answer a large model in the cloud does better.
A start-up wants to try out many different models quickly without buying hardware.
Without hardware of its own there is nothing to run models on locally; cloud services are simpler here.
A marketing team needs current competitor prices from the web.
A local model only knows what lies in its parameters; current prices need a web search.
What does it mean to quantize a model?
Storing the parameters with fewer bits, which makes the file smaller and costs a little quality.
Correct: the numbers are rounded so that the model fits into memory.
Removing rarely used words from the vocabulary so that the model gets faster.
The vocabulary stays the same; it is the numbers that are quantized, not the words.
Splitting the model into parts that run on several computers.
That would be distributing the model; the number of bits per parameter would stay the same.
Training the model a second time on less data.
Quantization changes a finished model without training it again.
Download with answers:
CSV
·
Anki
More on this
Details
Level Beginner
Modules Applications of Artificial Intelligence
Code 2148
Updated 4 November 2026
Status published