Ollama models: download size, VRAM and system requirements
Pick your GPU or Mac to see which Ollama models run on it. Each row is an Ollama library tag we matched to its original model, so the memory figures use that model's real architecture: download size plus KV cache plus runtime overhead. For example, qwen3:8b is a 5.2GB download and needs about 6.1 GB at Ollama's 4K default context.
70 models. Sizes as listed on ollama.com; memory is our estimate.
| Model | Command | Download | Memory at 4K ctx | at 32K ctx | Smallest GPU | Smallest Mac | On your hardware |
|---|---|---|---|---|---|---|---|
| Qwen3 0.6B 752M | ollama run qwen3:0.6b | 0.5 GB Q4_K_M | 1.6 GB | 4.7 GB | Arc B580 | M4 16GB | |
| Qwen3 14B 15B | ollama run qwen3:14b | 9.3 GB Q4_K_M | 10.0 GB | 14.4 GB | Arc B580 | M4 16GB | |
| Qwen3.8 27B 28B · vision | ollama run qwen3.8:27b | 18.0 GB Q4_K_M | 17.9 GB | 19.6 GB | RX 7900 XT | M1 Pro 32GB | |
| Qwen3 VL 8B Instruct 8.8B · vision | ollama run qwen3-vl:8b-instruct | 6.2 GB Q4_K_M | 7.0 GB | 11.0 GB | Arc B580 | M4 16GB | |
| Qwen2.5 7B Instruct 7.6B | ollama run qwen2.5:7b | 4.7 GB Q4_K_M | 5.3 GB | 6.8 GB | Arc B580 | M4 16GB | |
| Gemma 4 31B IT 31B · vision | ollama run gemma4:31b | 19.0 GB Q4_K_M | 20.0 GB | 24.4 GB | RX 7900 XTX | M1 Pro 32GB | |
| Qwen3.5 9B 9.7B · vision | ollama run qwen3.5:9b | 7.6 GB Q4_K_M | 8.1 GB | 9.0 GB | Arc B580 | M4 16GB | |
| Qwen3 8B 8.2B | ollama run qwen3:8b | 5.2 GB Q4_K_M | 6.1 GB | 10.0 GB | Arc B580 | M4 16GB | |
| gpt-oss 20B 21B | ollama run gpt-oss:20b | 14.0 GB MXFP4 | 13.9 GB | 14.6 GB | Arc A770 16GB | M4 Pro 24GB | |
| Qwen3.5 2B 2.3B · vision | ollama run qwen3.5:2b | 3.1 GB Q8_0 | 3.8 GB | 4.1 GB | Arc B580 | M4 16GB | |
| Qwen2.5 VL 7B Instruct 8.3B · vision | ollama run qwen2.5vl:7b | 6.0 GB Q4_K_M | 6.5 GB | 8.0 GB | Arc B580 | M4 16GB | |
| Gemma 4 26B A4B IT 26B · vision | ollama run gemma4:26b | 16.0 GB Q4_K_M | 16.2 GB | 17.2 GB | RX 7900 XT | M1 Pro 32GB | |
| Qwen2.5 1.5B Instruct 1.5B | ollama run qwen2.5:1.5b | 1.0 GB Q4_K_M | 1.7 GB | 2.5 GB | Arc B580 | M4 16GB | |
| Qwen3.5 4B 4.7B · vision | ollama run qwen3.5:4b | 4.0 GB Q4_K_M | 4.7 GB | 5.6 GB | Arc B580 | M4 16GB | |
| Llama 3.1 8B Instruct 8.0B | ollama run llama3.1:8b | 4.9 GB Q4_K_M | 5.7 GB | 9.2 GB | Arc B580 | M4 16GB | |
| Qwen2.5 0.5B Instruct 494M | ollama run qwen2.5:0.5b | 0.4 GB Q4_K_M | 1.1 GB | 1.4 GB | Arc B580 | M4 16GB | |
| Qwen3.6 35B A3B 36B · vision | ollama run qwen3.6:35b | 23.0 GB Q4_K_M | 22.4 GB | 22.9 GB | RX 7900 XTX | M3 Pro 36GB | |
| Qwen3 32B 33B | ollama run qwen3:32b | 20.0 GB Q4_K_M | 20.3 GB | 27.3 GB | RX 7900 XTX | M1 Pro 32GB | |
| Qwen3 VL 4B Instruct 4.4B · vision | ollama run qwen3-vl:4b-instruct | 3.3 GB Q4_K_M | 4.3 GB | 8.3 GB | Arc B580 | M4 16GB | |
| Qwen3.5 0.8B 873M · vision | ollama run qwen3.5:0.8b | 1.2 GB Q8_0 | 2.0 GB | 2.4 GB | Arc B580 | M4 16GB | |
| Qwen3 1.7B 2.0B | ollama run qwen3:1.7b | 1.4 GB Q4_K_M | 2.4 GB | 5.5 GB | Arc B580 | M4 16GB | |
| Qwen3 4B Instruct 2507 4.0B | ollama run qwen3:4b-instruct-2507-q4_K_M | 2.5 GB Q4_K_M | 3.6 GB | 7.5 GB | Arc B580 | M4 16GB | |
| Qwen3.6 27B 28B · vision | ollama run qwen3.6:27b | 19.0 GB Q4_K_M | 18.8 GB | 20.6 GB | RX 7900 XT | M1 Pro 32GB | |
| gpt-oss 120B 117B | ollama run gpt-oss:120b | 65.0 GB MXFP4 | 61.5 GB | 62.4 GB | none single | M2 Max 96GB | |
| Qwen3 Coder 30B A3B Instruct 31B | ollama run qwen3-coder:30b | 19.0 GB Q4_K_M | 18.8 GB | 21.4 GB | RX 7900 XT | M1 Pro 32GB | |
| Qwen2.5 3B Instruct 3.1B | ollama run qwen2.5:3b | 1.9 GB Q4_K_M | 2.6 GB | 3.6 GB | Arc B580 | M4 16GB | |
| Qwen2.5 VL 3B Instruct 3.8B · vision | ollama run qwen2.5vl:3b | 3.2 GB Q4_K_M | 3.8 GB | 4.8 GB | Arc B580 | M4 16GB | |
| DeepSeek OCR 3.3B | ollama run deepseek-ocr:3b | 6.7 GB F16 | 7.1 GB | max 8K | Arc B580 | M4 16GB | |
| Qwen3.5 27B 28B · vision | ollama run qwen3.5:27b | 20.0 GB Q4_K_M | 19.8 GB | 21.5 GB | RX 7900 XTX | M1 Pro 32GB | |
| Qwen2.5 Coder 14B Instruct 15B | ollama run qwen2.5-coder:14b | 9.0 GB Q4_K_M | 9.8 GB | 15.1 GB | Arc B580 | M4 16GB | |
| Qwen2.5 Coder 7B Instruct 7.6B | ollama run qwen2.5-coder:7b | 4.7 GB Q4_K_M | 5.3 GB | 6.8 GB | Arc B580 | M4 16GB | |
| Qwen3.5 35B A3B 36B · vision | ollama run qwen3.5:35b | 22.0 GB Q4_K_M | 21.4 GB | 22.0 GB | RX 7900 XTX | M3 Pro 36GB | |
| DeepSeek V3 684B | ollama run deepseek-v3:671b | 404 GB Q4_K_M | 377 GB | 379 GB | none single | M3 Ultra 512GB | |
| Llama 3.2 1B Instruct 1.2B | ollama run llama3.2:1b | 1.3 GB Q8_0 | 2.0 GB | 2.9 GB | Arc B580 | M4 16GB | |
| Mistral 7B Instruct v0.2 7.2B | ollama run mistral:7b-instruct-v0.2-q4_K_M | 4.4 GB Q4_K_M | 5.1 GB | 8.6 GB | Arc B580 | M4 16GB | |
| Gemma 3 4B IT 4.3B · vision | ollama run gemma3:4b | 3.4 GB Q4_K_M | 4.3 GB | 4.8 GB | Arc B580 | M4 16GB | |
| Llama 3.2 3B Instruct 3.2B | ollama run llama3.2:3b | 2.0 GB Q4_K_M | 2.9 GB | 6.0 GB | Arc B580 | M4 16GB | |
| Llama 3.2 1B 1.2B | ollama run llama3.2:1b-text-q4_K_M | 0.8 GB Q4_K_M | 1.5 GB | 2.4 GB | Arc B580 | M4 16GB | |
| DeepSeek R1 0528 Qwen3 8B 8.2B | ollama run deepseek-r1:8b | 5.2 GB Q4_K_M | 6.1 GB | 10.0 GB | Arc B580 | M4 16GB | |
| MedGemma 4B IT 4.3B · vision | ollama run medgemma:4b | 3.3 GB Q4_K_M | 4.2 GB | 4.7 GB | Arc B580 | M4 16GB | |
| DeepSeek R1 Distill Qwen 1.5B 1.8B | ollama run deepseek-r1:1.5b | 1.1 GB Q4_K_M | 1.8 GB | 2.6 GB | Arc B580 | M4 16GB | |
| Llama 3.1 8B 8.0B | ollama run llama3.1:8b-text-q4_K_M | 4.9 GB Q4_K_M | 5.7 GB | 9.2 GB | Arc B580 | M4 16GB | |
| Gemma 3 12B IT 12B · vision | ollama run gemma3:12b | 8.2 GB Q4_K_M | 9.1 GB | 10.8 GB | Arc B580 | M4 16GB | |
| DeepSeek R1 Distill Qwen 32B 33B | ollama run deepseek-r1:32b | 20.0 GB Q4_K_M | 20.3 GB | 27.3 GB | RX 7900 XTX | M1 Pro 32GB | |
| Llama 3.3 70B Instruct 71B | ollama run llama3.3:70b | 43.0 GB Q4_K_M | 41.9 GB | 50.7 GB | none single | M4 Max 64GB | |
| Hermes 3 Llama 3.1 8B 8.0B | ollama run hermes3:8b | 4.7 GB Q4_0 | 5.5 GB | 9.0 GB | Arc B580 | M4 16GB | |
| Phi 3.5 Mini Instruct 3.8B | ollama run phi3.5:3.8b | 2.2 GB Q4_0 | 4.0 GB | 14.5 GB | Arc B580 | M4 16GB | |
| Llama 3.2 3B 3.2B | ollama run llama3.2:3b-text-q4_K_M | 2.0 GB Q4_K_M | 2.9 GB | 6.0 GB | Arc B580 | M4 16GB | |
| Phi 3 Mini 4K Instruct 3.8B | ollama run phi3:3.8b-mini-4k-instruct-q4_K_M | 2.4 GB Q4_K_M | 4.2 GB | max 4K | Arc B580 | M4 16GB | |
| Gemma 3 27B IT 27B · vision | ollama run gemma3:27b | 17.0 GB Q4_K_M | 17.5 GB | 19.6 GB | RX 7900 XT | M1 Pro 32GB | |
| Phi 4 Mini Instruct 3.8B | ollama run phi4-mini:3.8b | 2.5 GB Q4_K_M | 3.6 GB | 7.1 GB | Arc B580 | M4 16GB | |
| DeepSeek R1 Distill Qwen 7B 7.6B | ollama run deepseek-r1:7b | 4.7 GB Q4_K_M | 5.3 GB | 6.8 GB | Arc B580 | M4 16GB | |
| SmolLM2 1.7B Instruct 1.7B | ollama run smollm2:1.7b | 1.8 GB Q8_0 | 2.9 GB | max 8K | Arc B580 | M4 16GB | |
| Phi 3 Mini 128K Instruct 3.8B | ollama run phi3:3.8b-mini-128k-instruct-q4_K_M | 2.4 GB Q4_K_M | 4.2 GB | 14.7 GB | Arc B580 | M4 16GB | |
| SmolLM2 360M Instruct 362M | ollama run smollm2:360m | 0.7 GB F16 | 1.3 GB | max 8K | Arc B580 | M4 16GB | |
| Phi 4 15B | ollama run phi4:14b | 9.1 GB Q4_K_M | 9.8 GB | max 16K | Arc B580 | M4 16GB | |
| DeepSeek R1 0528 685B | ollama run deepseek-r1:671b | 404 GB Q4_K_M | 377 GB | 379 GB | none single | M3 Ultra 512GB | |
| Meta Llama 3.1 8B Instruct 8.0B | ollama run llama3.1:8b | 4.9 GB Q4_K_M | 5.7 GB | 9.2 GB | Arc B580 | M4 16GB | |
| Granite 3.1 3B A800M Instruct 3.3B | ollama run granite3.1-moe:3b | 2.0 GB Q4_K_M | 2.6 GB | 4.4 GB | Arc B580 | M4 16GB | |
| DeepSeek R1 Distill Qwen 14B 15B | ollama run deepseek-r1:14b | 9.0 GB Q4_K_M | 9.8 GB | 15.1 GB | Arc B580 | M4 16GB | |
| Llama 4 Scout 17B 16E Instruct 109B | ollama run llama4:scout | 67.0 GB Q4_K_M | 63.9 GB | 69.2 GB | none single | M2 Max 96GB | |
| Granite 3.3 8B Instruct 8.2B | ollama run granite3.3:8b | 4.9 GB Q4_K_M | 5.7 GB | 10.1 GB | Arc B580 | M4 16GB | |
| Llama 3.1 70B Instruct 71B | ollama run llama3.1:70b | 43.0 GB Q4_K_M | 41.9 GB | 50.7 GB | none single | M4 Max 64GB | |
| DeepSeek Coder V2 Lite Instruct 16B | ollama run deepseek-coder-v2:16b | 8.9 GB Q4_0 | 9.0 GB | 9.8 GB | Arc B580 | M4 16GB | |
| Meta Llama 3 8B Instruct 8.0B | ollama run llama3:8b | 4.7 GB Q4_0 | 5.5 GB | max 8K | Arc B580 | M4 16GB | |
| DeepSeek R1 Distill Llama 8B 8.0B | ollama run deepseek-r1:8b-llama-distill-q4_K_M | 4.9 GB Q4_K_M | 5.7 GB | 9.2 GB | Arc B580 | M4 16GB | |
| Falcon 7B Instruct 7.2B | ollama run falcon:7b-instruct | 4.2 GB Q4_0 | 4.5 GB | 4.7 GB | Arc B580 | M4 16GB | |
| DeepSeek R1 Distill Llama 70B 71B | ollama run deepseek-r1:70b | 43.0 GB Q4_K_M | 41.9 GB | 50.7 GB | none single | M4 Max 64GB | |
| GLM 4 9B 9.4B | ollama run glm4:9b-text-q4_K_M | 6.3 GB Q4_K_M | 6.7 GB | max 8K | Arc B580 | M4 16GB | |
| Llama 4 Maverick 17B 128E Instruct 402B | ollama run llama4:maverick | 245 GB Q4_K_M | 230 GB | 235 GB | none single | M3 Ultra 512GB |
Download sizes are decimal GB, as ollama.com lists them; memory is binary GB (GiB) like the rest of this site. "Smallest GPU" is the cheapest consumer card (launch MSRP) that holds the model at 4K context. A tag is listed only when its architecture, parameter count and file size match the source model; others are left out rather than guessed. Data is collected automatically from ollama.com/library and refreshed weekly (last check Nov 17, 2026).
How much memory do Ollama models need?
Most Ollama tags are 4-bit (Q4_K_M) builds, so memory scales with parameter count. Medians across the 70 tags we match:
| Model size | Tags | Download | Memory at 4K ctx | Memory at 32K ctx | Examples |
|---|---|---|---|---|---|
| Up to 2B | 11 | 1.1 GB | 1.8 GB | 2.6 GB | qwen3:0.6b, qwen3.5:2b, qwen2.5:1.5b |
| 3–4B | 15 | 2.5 GB | 4.0 GB | 6.0 GB | qwen3.5:4b, qwen3-vl:4b-instruct, qwen3:4b-instruct-2507-q4_K_M |
| 7–9B | 18 | 4.9 GB | 5.7 GB | 9.0 GB | qwen3-vl:8b-instruct, qwen2.5:7b, qwen3.5:9b |
| 12–15B | 6 | 9.0 GB | 9.8 GB | 12.2 GB | qwen3:14b, qwen2.5-coder:14b, gemma3:12b |
| 20–32B | 12 | 19.0 GB | 18.8 GB | 21.4 GB | qwen3.8:27b, gemma4:31b, gpt-oss:20b |
| 70B | 3 | 43.0 GB | 41.9 GB | 50.7 GB | llama3.3:70b, llama3.1:70b, deepseek-r1:70b |
| 100B and up | 5 | 245 GB | 230 GB | 235 GB | gpt-oss:120b, deepseek-v3:671b, deepseek-r1:671b |
As a rule of thumb, an 8 GB GPU handles the 7–9B class at 4K context, 12–16 GB adds the 12–15B class and longer contexts, and 24 GB is where 20–32B models become comfortable. Larger models need a Mac with lots of unified memory, several GPUs, or partial CPU offload, which Ollama does automatically but which is much slower. Browse by size on models by VRAM.
Ollama system requirements
| macOS | macOS 14 Sonoma or newer. Apple Silicon (M1 and later) runs models on the GPU through Metal; Intel Macs run on the CPU only. |
| Windows | Windows 10 22H2 or newer. NVIDIA driver 551.61 or newer for NVIDIA cards; for AMD, a ROCm v7 driver (RX 7600 to 7900 XTX) or a Vulkan-capable Radeon driver. |
| Linux | NVIDIA: compute capability 5.0+ (GTX 750 and newer) with driver 550+. AMD: ROCm v7 for RX 6800 and newer, RX 9000 series and Ryzen AI Max. Other GPUs, including Intel Arc, via Vulkan. |
| Disk | About 4 GB for the program on Windows, plus the models: from under 1 GB to hundreds of GB each (see the download column above). |
| Memory | The model must fit in GPU memory to run at full speed. If it doesn't, Ollama splits it between GPU and system RAM, which runs at a fraction of the speed. Use the filter above, or the VRAM calculator for other quants and context lengths. |
From Ollama's documentation, checked Nov 17, 2026: hardware support, macOS, Windows. Check there for changes.
Why Ollama uses more memory on bigger GPUs
Ollama picks a default context length from your GPU memory: 4K tokens below 24 GB, 32K from 24 to 48 GB, and 256K at 48 GB and above (docs). A longer context needs a bigger KV cache, so the same model can take several GB more on a 24 GB card than on a 16 GB one. The filter above uses that default for your hardware. To change it, set OLLAMA_CONTEXT_LENGTH (for example OLLAMA_CONTEXT_LENGTH=8192 ollama serve) or move the slider in the app, then check the split with ollama ps.
Questions
How much VRAM do I need for Ollama?
It depends on the model, not on Ollama. A 4-bit 7–9B model needs about 5.7 GB at the default 4K context, so an 8 GB GPU is a practical minimum for useful models; 1–4B models run on less. Each model's page has the exact figure by quant and context.
Does Ollama use the same files as LM Studio and llama.cpp?
All three run GGUF models with llama.cpp-derived code, so memory needs are close. Ollama's default tags are usually Q4_K_M; some are Q4_0 or Q8_0, shown in the download column. See the LM Studio guide and the llama.cpp offload guide.
Which Ollama model should I start with?
Pick the largest model that runs fully on your hardware. Use the filter above, or read the beginner's guide to running an LLM locally.
Automated data: model architecture from the Hugging Face Hub API, Ollama tags and sizes from ollama.com, formulas on the methodology page. Not affiliated with Ollama. Report an error.