Raphael Friedmann
← Understanding AI

How to read a model's name and specs

Model names like Qwen3-30B-A3B-Q4_K_M look like a cat walked across the keyboard. Here's how to decode them, and the handful of specs that actually matter.

Go looking for a model to download and you immediately hit names like Qwen3-30B-A3B-Q4_K_M, and specs like “128k context, 300 tok/s”. It reads like a cat walked across the keyboard. But every piece of it means something useful, and once you can read it you can size up any model at a glance.

The name

A model’s name is really a stack of labels. Here’s a typical one, taken apart:

Reading a model's name
Qwen3modelfamily30B30 billionparametersA3B3B activeper tokenQ4_K_M4-bitshrunk to fit
Once you know the parts, any model name on a site like Hugging Face tells you its size, its speed trick, and how much memory it needs.

Going left to right:

The A3B part is a twist some models have. They’re built from many “experts” and only switch on a slice of themselves for each word, so a model can be huge in total but run as quickly as a small one. “30B total, 3B active” means exactly that: big brain, light on its feet.

The Q4 part is the quantisation: the model has been shrunk to 4-bit numbers so it fits on a normal card. The extra letters (K_M and friends) are just flavours of how that shrinking was done.

The two specs worth knowing

Beyond the name, two numbers tell you how it’ll feel to use.

That’s the whole vocabulary. Size (parameters), speed trick (active params), how it was shrunk (quantization), how much it can hold (context), and how fast it writes (tok/s). With those five, a wall of cryptic model names turns into a menu you can actually order from.

From here, the natural next questions are what these models even are and why you’d run one yourself.