The Serving Desk
Each level is a customer who needs a model served. First predict one number, then choose the GPU, the number format and how many copies to run. Your plan is checked with the same maths the course teaches, and you score only for being right.
Level 1: Will the model fit on one card?
A model's weights must fit in the GPU's memory before anything else can happen, and their size is the number of parameters times the bytes each one takes.
- Customer
- A town library
- Model
- Qwen3 8B
- People at once
- 1
- Tokens each
- 1,024 in, 512 out
- First word within
- 1.00 s
- Each word within
- 50 ms
- Budget
- $1.00 an hour
What the customer said
We want a chat helper that answers questions about opening hours and lending rules. One reader at a time is enough for now, and we have chosen Qwen3 8B.