What does 12B mean in AI models?

What a parameter is

A language model is a very large network of numbers. During training, each number is nudged billions of times until the network gets good at predicting the next token. Those settled numbers are the parameters. When the model writes, every new token is computed through them, so the count is a rough measure of how much the model can hold and how much work each word takes.

Model names pack several facts together. "Gemma 4 12B" is the Gemma family, fourth generation, about 12 billion parameters. Other parts you will meet:

• "Instruct" or "it": tuned to follow chat instructions, the version roleplay apps use.
• "Base": the raw model before chat tuning.
• Two sizes, like "30B-A3B": a mixture-of-experts model with 30 billion parameters in total but only about 3 billion active per token. It writes fast like a small model but needs memory like a big one.
• "Q4", "Q8", "FP8": the model has been quantized, stored with fewer bits per parameter.

Size and memory, roughly

The weights must sit in memory while the model runs. A rough rule is bytes per parameter times the count:

• 16-bit: 2 bytes each, so 12B needs about 24 GB and 31B about 62 GB.
• 8-bit: about half, so roughly 12 GB and 31 GB.
• 4-bit: about a quarter, so roughly 7 GB and 17 GB.

On top of that comes memory for the context, which grows as the chat gets longer. That is why people running models at home choose a size by their graphics card, and why a 12B is a popular home size while a 31B usually needs a large card or a hosted service.

What size changes in a story

Larger models tend to keep several characters distinct, read subtext, hold a secret for fifty messages and follow complicated cards. Smaller ones answer faster and are often all a one-on-one scene needs. Returns shrink as size grows, and training matters as much as count: a newer, well-tuned model can write better than an older, larger one.

Size is not context. Parameter count says nothing about how much of the story a model can read at once; on UnsavedAI, Gemma 4 12B and 31B both read 61,440 tokens including the reply.

Common mistakes

• Reading B as bytes or bits. It is billions of parameters.
• Assuming bigger is always better for you. Speed matters in roleplay, and a reply that starts quickly keeps a scene alive.
• Comparing sizes across families and years as if they were the same scale.
• Treating the total of a mixture-of-experts model like a dense model of that size.
• Thinking you need your own GPU to use a 12B. Hosted APIs run it for you; see SillyTavern without a GPU.

12B and 31B on UnsavedAI

Every UnsavedAI plan includes gemma-4-12b, described on Models as quick and lively; Priority adds gemma-4-31b, deep and thoughtful. Both power every house character on the site and both work through the API. The simplest test is to play the same opening twice, once on each, with a character who juggles a lot, like Queen Sylvara and her court. For a longer comparison and the settings that suit each, read Gemma 4 for roleplay.

Questions people ask