Local LLM VRAM calculator: which free AI model fits your GPU, Mac or PC?
Model sizes checked November 11, 2026 on the Ollama library. Models get re-uploaded and new tags appear every few weeks, so we recheck monthly.
Short answer: a model needs roughly its download size in memory, plus room for the conversation. At Ollama's default downloads, 8 GB of VRAM runs the 8B class (Qwen3 8B, Llama 3.1 8B, DeepSeek-R1 8B). 12 GB adds the 14B class at about 9 GB. 16 GB gets you gpt-oss 20B (14 GB) and Mistral Small 3.2 (15 GB). 24 GB is where the 27B to 32B models live, at 17 to 20 GB each. Anything 70B and up (43 GB or more) needs two cards, a big Mac, or patience with CPU offload.
The calculator below does this for your exact hardware.
The calculator
The calculator needs JavaScript. Without it: memory needed = download size + about 1.5 GB. If that fits in your VRAM, the model runs fully on the GPU. The tables below cover common setups.
How it decides. A model "fits" when its download size plus your headroom is no more than your VRAM. If it doesn't fit but would fit in VRAM plus system RAM (minus 6 GB we keep back for the operating system), Ollama can still run it by putting some layers on the CPU, which works but is much slower. On a Mac, the GPU shares memory with everything else, so we count 70% of unified memory as usable. That percentage is our working assumption, not an Apple figure, so change it if you know better for your machine.
What fits on common GPUs
Same math as the calculator: 1.5 GB headroom, 32 GB of system RAM. The last column names the three biggest models that stay 100% on the card.
| VRAM | Models that fit fully | Biggest that fit |
|---|---|---|
| 6 GB | 5 of 33 | Qwen3.5 4B (4 GB), Qwen3.5 2B (3.1 GB), Ministral 3 3B (3 GB) |
| 8 GB | 9 of 33 | Ministral 3 8B (6 GB), DeepSeek-R1 8B (Qwen3 distill) (5.2 GB), Qwen3 8B (5.2 GB) |
| 12 GB | 17 of 33 | Gemma 4 E4B (9.5 GB), Qwen3 14B (9.3 GB), Ministral 3 14B (9.1 GB) |
| 16 GB | 18 of 33 | gpt-oss 20B (14 GB), Gemma 4 E4B (9.5 GB), Qwen3 14B (9.3 GB) |
| 24 GB | 27 of 33 | Qwen3.5 35B (22 GB), DeepSeek-R1 32B (20 GB), Qwen3.5 27B (20 GB) |
| 32 GB | 28 of 33 | Qwen3.6 35B-A3B (MoE) (23 GB), Qwen3.5 35B (22 GB), DeepSeek-R1 32B (20 GB) |
What fits on Apple silicon Macs
Assumes 70% of unified memory is usable for the model.
| Unified memory | Models that fit | Biggest that fit |
|---|---|---|
| 16 GB | 17 of 33 | Gemma 4 E4B (9.5 GB), Qwen3 14B (9.3 GB), Ministral 3 14B (9.1 GB) |
| 24 GB | 19 of 33 | Mistral Small 3.2 24B (15 GB), gpt-oss 20B (14 GB), Gemma 4 E4B (9.5 GB) |
| 32 GB | 26 of 33 | DeepSeek-R1 32B (20 GB), Qwen3.5 27B (20 GB), Gemma 4 31B (20 GB) |
| 48 GB | 28 of 33 | Qwen3.6 35B-A3B (MoE) (23 GB), Qwen3.5 35B (22 GB), DeepSeek-R1 32B (20 GB) |
| 64 GB | 30 of 33 | DeepSeek-R1 70B (Llama distill) (43 GB), Llama 3.1 70B (43 GB), Qwen3.6 35B-A3B (MoE) (23 GB) |
| 128 GB | 33 of 33 | Qwen3.5 122B (81 GB), Llama 4 Scout (67 GB), gpt-oss 120B (65 GB) |
Every model in the calculator
Download = what Ollama's library page shows for that tag on the day we checked. Mostly 4-bit Q4_K_M files. Qwen3.5 9B is listed as 6.6 to 7.6 GB, and a couple of Gemma 4 tags are similar, so where there's a range we've used the top end. Better to be told a model won't fit and find it does.
| Model | Ollama tag | Download | Memory to plan for | Source |
|---|---|---|---|---|
| Qwen3.5 0.8B | qwen3.5:0.8b |
1.3 GB | 2.8 GB | page |
| Qwen3 4B | qwen3:4b |
2.5 GB | 4.0 GB | page |
| Ministral 3 3B | ministral-3:3b |
3 GB | 4.5 GB | page |
| Qwen3.5 2B | qwen3.5:2b |
3.1 GB | 4.6 GB | page |
| Qwen3.5 4B | qwen3.5:4b |
4 GB | 5.5 GB | page |
| Llama 3.1 8B | llama3.1:8b |
4.9 GB | 6.4 GB | page |
| Qwen3 8B | qwen3:8b |
5.2 GB | 6.7 GB | page |
| DeepSeek-R1 8B (Qwen3 distill) | deepseek-r1:8b |
5.2 GB | 6.7 GB | page |
| Ministral 3 8B | ministral-3:8b |
6 GB | 7.5 GB | page |
| Gemma 4 E2B | gemma4:e2b |
7.5 GB | 9.0 GB | page |
| Qwen3.5 9B | qwen3.5:9b |
7.6 GB | 9.1 GB | page |
| Gemma 4 12B | gemma4:12b |
8 GB | 9.5 GB | page |
| DeepSeek-R1 14B | deepseek-r1:14b |
9 GB | 10.5 GB | page |
| Phi-4 14B | phi4:14b |
9.1 GB | 10.6 GB | page |
| Ministral 3 14B | ministral-3:14b |
9.1 GB | 10.6 GB | page |
| Qwen3 14B | qwen3:14b |
9.3 GB | 10.8 GB | page |
| Gemma 4 E4B | gemma4:e4b |
9.5 GB | 11.0 GB | page |
| gpt-oss 20B | gpt-oss:20b |
14 GB | 15.5 GB | page |
| Mistral Small 3.2 24B | mistral-small3.2:24b |
15 GB | 16.5 GB | page |
| Qwen3.6 27B | qwen3.6:27b |
18 GB | 19.5 GB | page |
| Gemma 4 26B (MoE, 3.8B active) | gemma4:26b |
19 GB | 20.5 GB | page |
| Qwen3 30B-A3B (MoE) | qwen3:30b |
19 GB | 20.5 GB | page |
| GLM-4.7-Flash | glm-4.7-flash |
19 GB | 20.5 GB | page |
| Gemma 4 31B | gemma4:31b |
20 GB | 21.5 GB | page |
| Qwen3.5 27B | qwen3.5:27b |
20 GB | 21.5 GB | page |
| DeepSeek-R1 32B | deepseek-r1:32b |
20 GB | 21.5 GB | page |
| Qwen3.5 35B | qwen3.5:35b |
22 GB | 23.5 GB | page |
| Qwen3.6 35B-A3B (MoE) | qwen3.6:35b-a3b |
23 GB | 24.5 GB | page |
| Llama 3.1 70B | llama3.1:70b |
43 GB | 44.5 GB | page |
| DeepSeek-R1 70B (Llama distill) | deepseek-r1:70b |
43 GB | 44.5 GB | page |
| gpt-oss 120B | gpt-oss:120b |
65 GB | 66.5 GB | page |
| Llama 4 Scout | llama4:scout |
67 GB | 68.5 GB | page |
| Qwen3.5 122B | qwen3.5:122b |
81 GB | 82.5 GB | page |
Download this table as a CSV file. Free to reuse with a link back.
Why your real number can differ
- Context length. Ollama's context length docs say a longer context needs more memory. By default Ollama uses 4k context below 24 GiB of VRAM, 32k from 24 to 48 GiB, and 256k at 48 GiB and up. If you raise it for coding tools or long documents, add more headroom.
- Other apps. Your browser, a game launcher and the desktop itself all take VRAM. On a laptop with one GPU driving the screen, assume a little less than the box says.
- Image input. Many of these models also read images. The vision part adds to the download size, which is part of why some tags show a size range.
- Mixture of experts. Models like Gemma 4 26B only use part of their weights for each word. Ollama's page lists 3.8B active out of 25.2B total. They still need all the weights in memory, but they tend to cope better when part of the model sits in system RAM.
Check it for real. Load a model, then run ollama ps. The PROCESSOR column shows "100% GPU" when it fits, or a GPU and CPU split when it doesn't. That one command settles any argument with a calculator, including this one.
Is your GPU supported?
Ollama's hardware support page lists NVIDIA cards with compute capability 5.0 or newer and driver 550 or newer, which covers GeForce GTX 750 Ti and everything since, through the RTX 50 series. AMD Radeon support on Linux and Windows needs a ROCm v7 driver stack, with Vulkan as the fallback for cards ROCm doesn't cover. On Windows, Ollama needs Windows 10 22H2 or newer and NVIDIA driver 551.61 or newer (Windows docs). On a Mac it needs macOS Sonoma (14) or newer, and only Apple silicon gets GPU acceleration (macOS docs).
Sources
Checked November 11, 2026.
- Ollama library: qwen3.5 (tags and download sizes)
- Ollama library: qwen3 (tags and download sizes)
- Ollama library: ministral-3 (tags and download sizes)
- Ollama library: llama3.1 (tags and download sizes)
- Ollama library: deepseek-r1 (tags and download sizes)
- Ollama library: gemma4 (tags and download sizes)
- Ollama library: phi4 (tags and download sizes)
- Ollama library: gpt-oss (tags and download sizes)
- Ollama library: mistral-small3.2 (tags and download sizes)
- Ollama library: qwen3.6 (tags and download sizes)
- Ollama library: glm-4.7-flash (tags and download sizes)
- Ollama library: llama4 (tags and download sizes)
- Ollama docs: Context length, Hardware support, Windows, macOS
Spot a size that's changed? Tell us and we'll update the table.
Quick answers
How much VRAM do I need to run a local LLM?
Roughly the model's download size plus about 1.5 GB for the context and runtime. At Ollama's default 4-bit downloads, an 8B model like Llama 3.1 8B is 4.9 GB, so 6 to 8 GB of VRAM is enough. A 14B model is about 9 GB and wants 12 GB of VRAM. Most 27B to 32B models are 17 to 20 GB and need a 24 GB card.
What is the best model for an 8 GB GPU?
On an 8 GB card, models up to about 6.5 GB fit fully in VRAM with our headroom assumption, which includes Qwen3 8B (5.2 GB), Llama 3.1 8B (4.9 GB), DeepSeek-R1 8B (5.2 GB) and Ministral 3 8B (6.0 GB). Larger models will still run if you have enough system RAM, but Ollama splits them between GPU and CPU and they slow down a lot.
Can a 16 GB Mac run local AI models?
Yes, small and mid-size ones. With our assumption that about 70% of unified memory is usable for the model, a 16 GB Mac has roughly 11 GB to work with, enough for 8B to 14B models such as Qwen3 14B (9.3 GB). A 24 GB to 32 GB Mac is the comfortable starting point for 20B-class models like gpt-oss 20B (14 GB).
Does a longer context window need more memory?
Yes. Ollama's docs say a larger context length increases the memory a model needs. Ollama sets 4k context by default on GPUs with less than 24 GiB of VRAM, 32k on 24 to 48 GiB and 256k on 48 GiB or more. Run ollama ps to see how much memory a loaded model really uses and whether it is 100% on the GPU.