How big is a model?
What “7B” or “35B-A3B” means, and how size turns into gigabytes.
4 min · you’ll learn
- What parameters are
- How to estimate a model’s size
- Why MoE models are fast
“7B” means 7 billion parameters
A model is made of parameters: billions of numbers it learned during training. The “B” in a name like Qwen3.5 9B is how many billions it has.
More parameters usually means a smarter model, and always means a bigger, slower one.
Think of it like this
Parameters are like the knowledge in a library. A bigger library knows more, but it takes up more space and longer to search.
From parameters to gigabytes
At full precision every parameter takes 2 bytes. So a quick rule: full-size model ≈ parameters × 2 GB. Shrunk to the usual 4-bit version (next lesson), it’s roughly parameters × 0.6 GB.
Typical download size at 4-bit (Q4)
- 4B modelphones, any laptop≈3 GB
- 9B model8–16 GB computers≈6 GB
- 27B model32 GB+≈17 GB
- 70B model64 GB+≈42 GB
- 16 GB computer
- up to ~9B
- comfortably
- 32–48 GB
- ~27–32B
- the sweet spot
- 64 GB+
- 70B and up
- big models
Mixture of Experts: big, but fast
Some models are split into many smaller “experts”. For each word, a router picks just a few of them to do the work. These are called MoE models.
Memory holds all 16 experts.
Speed depends on the 2 that work per word.
Reading names like “35B-A3B”
MoE names show two numbers: 35B-A3B means 35 billion parameters in total, about 3 billion active for each word.
- Memory follows the total (35B): it needs as much room as any 35B model.
- Speed follows the active part (3B): it writes almost as fast as a tiny 3B model.
That makes MoE models a great match for Macs and other unified-memory machines: lots of room, moderate speed.
Remember this
Quick check
Which will usually write faster on the same computer?
Up next · lesson 4
Quantization: shrinking models
How a model gets 3–4× smaller with only a small loss in quality.