Ollama VRAM requirements: how much GPU memory each model needs

How much VRAM a model needs in Ollama, tag by tag. Each row uses the real model file size, quantization and context window from Ollama's registry for 39 popular tags, an estimate of total memory at 4k and 32k context, and the cheapest GPU and Mac we track that runs it fully on the GPU.

Tags

39 Ollama library defaults

Default quant

Q4_K_M for most tags

File sizes

2026-11-12 from registry.ollama.ai

Short answer: a Q4 Ollama model needs about its model file size plus 15% plus 1 GB. That means ~6.7 GB for an 8B tag, ~16.9 GB for gpt-oss:20b, ~24.2 GB for a 32B tag and ~49.9 GB for a 70B tag at Ollama's default 4k context. More context needs more memory.

How we estimate it

Same formula as the rest of the site (how we calculate), applied to the actual model file Ollama pulls instead of a bits-per-weight guess. These are estimates, not measured with ollama ps. "Smallest card" is the smallest common VRAM size where the model uses 90% or less. Cheapest GPU and Mac are the cheapest current devices we track that fit at 4k, at dated list prices. Mac memory is unified, and macOS keeps some of it, so we use the usable amount.

VRAM by Ollama tag

Model file size and quantization from each tag's manifest on registry.ollama.ai, context window from the tag list on ollama.com, retrieved 2026-11-12 (for example ollama.com/library/qwen3/tags). "Model file" is the language-model weights; Gemma tags also pull a vision projector (about 0.2–1 GB) that adds memory when you send images. Most default tags are Q4_K_M. llama3.2:1b defaults to Q8_0, mistral-nemo to Q4_0, and gpt-oss ships as MXFP4. "—" in the 32k column means we don't have the model's attention config yet; "max Nk ctx" means the tag's context window (from Ollama's tags page) is shorter than 32k. Tags can be updated by Ollama, so sizes may change.

Which Ollama models fit your GPU

8 GB VRAM

14 of these 39 tags fit. Largest: qwen3:8b (~7.0 GB), deepseek-r1:8b (~7.0 GB), granite3.3:8b (~6.7 GB), llama3.1:8b (~6.7 GB), deepseek-r1:7b (~6.4 GB), qwen2.5-coder:7b (~6.4 GB).

12 GB VRAM

19 of these 39 tags fit. Largest: olmo2:13b (~10.6 GB), gemma4:12b (~9.5 GB), gemma3:12b (~9.4 GB), mistral-nemo:12b (~9.1 GB), gemma4:e4b (~7.3 GB), qwen3:8b (~7.0 GB).

16 GB VRAM

23 of these 39 tags fit. Largest: qwen3:14b (~11.7 GB), phi4:14b (~11.4 GB), deepseek-r1:14b (~11.3 GB), qwen2.5-coder:14b (~11.3 GB), olmo2:13b (~10.6 GB), gemma4:12b (~9.5 GB).

24 GB VRAM

27 of these 39 tags fit. Largest: gemma4:26b (~20.6 GB), gemma3:27b (~20.0 GB), mistral-small3.2:24b (~18.5 GB), gpt-oss:20b (~16.9 GB, ✓32k), qwen3:14b (~11.7 GB, ✓32k), phi4:14b (~11.4 GB). Ollama defaults to 32k context on 24–48 GB cards, so ✓32k marks tags that still fit there.

32 GB VRAM

33 of these 39 tags fit. Largest: qwen3:32b (~24.2 GB), deepseek-r1:32b (~23.8 GB), qwen2.5-coder:32b (~23.8 GB), gemma4:31b (~22.5 GB), qwen3:30b (~22.3 GB, ✓32k), qwen3-coder:30b (~22.3 GB, ✓32k). Ollama defaults to 32k context on 24–48 GB cards, so ✓32k marks tags that still fit there.

48 GB VRAM

33 of these 39 tags fit. Largest: qwen3:32b (~24.2 GB), deepseek-r1:32b (~23.8 GB), qwen2.5-coder:32b (~23.8 GB), gemma4:31b (~22.5 GB), qwen3:30b (~22.3 GB), qwen3-coder:30b (~22.3 GB). Ollama defaults to 256k context on 48 GB+ cards, which can need far more memory than the 4k numbers here; lower it if ollama ps shows CPU offload.

64 GB VRAM

36 of these 39 tags fit. Largest: qwen2.5:72b (~55.5 GB), llama3.3:70b (~49.9 GB), deepseek-r1:70b (~49.9 GB), qwen3:32b (~24.2 GB), deepseek-r1:32b (~23.8 GB), qwen2.5-coder:32b (~23.8 GB). Ollama defaults to 256k context on 48 GB+ cards, which can need far more memory than the 4k numbers here; lower it if ollama ps shows CPU offload.

128 GB VRAM

37 of these 39 tags fit. Largest: gpt-oss:120b (~76.2 GB), qwen2.5:72b (~55.5 GB), llama3.3:70b (~49.9 GB), deepseek-r1:70b (~49.9 GB), qwen3:32b (~24.2 GB), deepseek-r1:32b (~23.8 GB). Ollama defaults to 256k context on 48 GB+ cards, which can need far more memory than the 4k numbers here; lower it if ollama ps shows CPU offload.

Ollama default context and why it matters

Ollama picks a default context length from your VRAM (docs): 4k under 24 GiB, 32k on 24–48 GiB, and 256k at 48 GiB or more. The KV cache grows with context, so a model that fits at 4k can spill to the CPU at 32k. For example, qwen3:30b needs about 22.3 GB at 4k and 25.2 GB at 32k. See context length and VRAM for the math. To set it yourself: OLLAMA_CONTEXT_LENGTH=8192 ollama serve, or /set parameter num_ctx 8192 inside ollama run.

Check the result with ollama ps. The PROCESSOR column should read 100% GPU. A split like "48%/52% CPU/GPU" means part of the model is on the CPU and speed drops a lot.

How much VRAM by model size · GGUF VRAM by quant (llama.cpp, LM Studio) · VRAM calculator · Q4 vs Q8 vs FP16 · Context length and VRAM · Tokens per second · Best GPU for local LLMs · Best LLMs by VRAM size