Skip to content
Local AI Wizard
Your machine
Theme

What is a model family?

One release line, several sizes, and many files for each size.

Models and families · updated

Short answer

One release line from one publisher, such as Qwen3.5, offered in several sizes.

  • Each size: A separate model, with its own scores and memory needs.
  • Files: Every size comes in many files: GGUF, MLX, Q4, Q8 and more.
  • In Qwen3.5: The 9B needs about 15 GB in Q4 on a 32 GB+ Mac; the 35B about 35 GB.

Planning estimate with the site's formula, not a measurement. Methodology

Three levels

Two of the Qwen3.5 sizes, the ones in this site's catalog. File sizes are the site's estimates for each precision.

Qwen3.5 family: sizes and file versionsSwipe sideways to see the files, or read the list below.
Qwen3.5 family: sizes and file versionsFamilySizesFile versionsQwen3.5AlibabaQwen3.5 9BGGUF Q4_K_M~5.6 GBGGUF Q8_0~10.1 GBBF16 original~18 GBQwen3.5 35B A3B3B activeGGUF Q4_K_M~21.9 GBGGUF Q8_0~39.4 GBBF16 original~70 GB

The same tree as a list:

  • Qwen3.5 · Alibaba
    • Qwen3.5 9B
      • GGUF Q4_K_M · ~5.6 GB
      • GGUF Q8_0 · ~10.1 GB
      • BF16 original · ~18 GB
    • Qwen3.5 35B A3B · 3B active
      • GGUF Q4_K_M · ~21.9 GB
      • GGUF Q8_0 · ~39.4 GB
      • BF16 original · ~70 GB

Sizes are separate models

Each size is trained on its own and tested on its own: a score for the large one says nothing about the small one. That is why the site gives every size its own page.

Q4 file, short conversations. Planning estimates in whole GB.
SizeTotal parametersActive per word32 GB+ Mac or PC (GB)Graphics card (GB)
Qwen3.5 9B9B9B159
Qwen3.5 35B A3B35B3B3529

Model cards: (Qwen/Qwen3.5-9B, reviewed Sep 30, 2026) · (Qwen/Qwen3.5-35B-A3B, reviewed Sep 30, 2026)

Reading a model name

Qwen3.5
Family and generationA new number is a new generation, trained again.
35B
Total parameters, in billionsSets the memory.
A3B
Active parameters per word (MoE models)Sets the work per word, not the memory.
Base · chat
Suffixes for how it was trained to answerA “Base” model only continues text; chat versions (often “Instruct”) follow requests. Each model card says which is which.

Choosing in a family

Choose the size first, by memory: the largest size whose estimate fits your machine is usually the best answer in that family. Then choose the file: a 9B in Q4 is about 5.6 GB of weights, the same model in BF16 about 18 GB (bytes per parameter).

Family pages

Every family with at least two open sizes in the catalog has its own page: sizes side by side, which one fits which machine, and the files published for each.