DOA
Help me choose

Local AI › Which local AI model to pick for your RAM: 8, 16, 24, 32 GB

Which local AI model to pick for your RAM: 8, 16, 24, 32 GB

Which model fits in your RAM (or your graphics card's VRAM), for chat and coding. Sizes are the downloaded file as published by Ollama and Hugging Face.

Checked on 13 Oct 2026. Versions change every week: check the official site before installing.

RAM/VRAMChatCoding
8 GB
  • qwen3.5:4b: ≈ 3.3 GB · Ollama
  • gemma4 (e4b): ≈ 6.6 GB (tight) · Ollama
  • qwen3.5:4b: ≈ 3.3 GB (general-purpose) · Ollama
16 GB
  • qwen3.5:9b: ≈ 6.6 GB · Ollama
  • Ornith-1.5-9B: Q4_K_M 5.8 GB · Q5_K_M 6.6 GB · Q8_0 9.8 GB · Hugging Face
  • gpt-oss:20b: ≈ 14 GB (tight) · Ollama
24 GB
  • Qwen3-Coder-30B-A3B: Q4_K_M 18.6 GB · Q5_K_M 21.7 GB · Hugging Face
  • qwen3-coder:30b: ≈ 19 GB · Ollama
32 GB+
  • Qwen3.8-27B: Q8_0 28.6 GB · Hugging Face
  • gpt-oss:120b: ≈ 65 GB (64 GB or more) · Ollama

Third-party data: file sizes are published by Ollama and Hugging Face on the linked pages; they are not DOA measurements.

How to read the table

  • The file has to fit in memory, and you need some extra room for the context and the operating system.
  • Q4, Q5 and Q8 (e.g. Q4_K_M) are quantisation levels: the higher the number, the bigger the file and the closer to the original model. Compare the sizes for the same model in the table.
  • MoE models («A3B») only activate part of their parameters, which makes them fast, but the whole file still has to fit in memory.
  • Models marked «-cloud» in the Ollama library do not run on your computer.
  • DOA does not recommend «uncensored» or «abliterated» variants.

And for images?

We do not give memory figures for image models because we could not verify them in an official source. See Generate AI images locally with ComfyUI.

Keep reading