Which AI model fits my graphics card?
Local AI is fastest when the whole model fits into video memory (VRAM). Choose your graphics card and the context length – the calculator shows which coding models fit completely.
All models at a glance
Estimated video memory needed per context length – calculated as described below. “Fits completely from” applies to a 16,000-token context.
| Model | File | KV cache per token | 8K context | 16K context | 32K context | Fits completely from |
|---|---|---|---|---|---|---|
| Qwen2.5-Coder 1.5B 1.5B parameters | 1.0 GB | 28 KB | 1.9 GB | 2.4 GB | 3.2 GB | 4 GB |
| Qwen2.5-Coder 3B 3B parameters | 2.0 GB | 36 KB | 3.0 GB | 3.5 GB | 4.6 GB | 6 GB |
| Qwen2.5-Coder 7B 7B parameters | 4.4 GB | 56 KB | 5.7 GB | 6.6 GB | 8.4 GB | 10 GB |
| Qwen2.5-Coder 14B 14B parameters | 8.4 GB | 192 KB | 11.0 GB | 13.1 GB | 17.4 GB | 16 GB |
| gpt-oss 20B MoE: 21B parameters, 3.6B active | 11.3 GB | 24 KB | 13.0 GB | 14.2 GB | 16.5 GB | 16 GB |
| Qwen3-Coder 30B-A3B MoE: 30.5B parameters, 3.3B active | 17.3 GB | 96 KB | 19.0 GB | 20.3 GB | 22.8 GB | 24 GB |
| Qwen2.5-Coder 32B 32B parameters | 18.5 GB | 256 KB | 21.6 GB | 24.2 GB | 29.5 GB | 32 GB |
How we calculate
Video memory needed = model file + KV cache + compute buffer + 0.5 GB for driver and runtime.
- Model file: the real size of the GGUF file VibeCoder Studio downloads (Q4_K_M, MXFP4 for gpt-oss).
- KV cache: the memory for the context. 24 to 256 KB per token depending on the model – calculated from the models' architecture data (layers, KV heads, head size) and checked against our measurements.
- Compute buffer: room for intermediate results while reading long inputs; it also grows with the context. We calculate conservatively – without the memory-saving flash attention, which is not active on every graphics card.
- Windows, the display and the browser use video memory themselves. “Fits completely” therefore means at least 1.5 GB of headroom; with at least 0.75 GB we show “tight”.
All figures are estimates as of November 2026. Sizes in GB as usual for graphics cards (1 GB = 1,024 MB).
Measured on our test PC
NVIDIA GeForce RTX 5080 (16 GB), AMD Ryzen 7 7800X3D, 32 GB RAM, Windows 11. The figures are medians of 18 to 46 replies to coding tasks.
| Model | Runtime | Context | Output (tokens/s) |
|---|---|---|---|
| Qwen2.5-Coder 7B | VibeCoder Studio (llama.cpp, Vulkan) | 32K | 133 |
| gpt-oss 20B | Ollama (CUDA) | 40K | 139 |
| Qwen3-Coder 30B-A3B | Ollama (CUDA), partly in system RAM | 16K | 69 |
Qwen3-Coder 30B-A3B does not fit completely into 16 GB. Because only 3.3B of its parameters are active per token, it stays fast even when partly offloaded. The Ollama runs used flash attention – with it, gpt-oss 20B needed only about 12.8 GB at 40K. The calculator above conservatively assumes no flash attention and therefore shows “Too large” from 32K.
Frequently asked questions
Where can I see how much video memory I have?
On Windows: open Task Manager (Ctrl + Shift + Esc), click “Performance” and then “GPU”. What counts is “Dedicated GPU memory”. “Shared GPU memory” is regular system RAM and does not count here.
Do I need an NVIDIA graphics card?
No. VibeCoder Studio uses llama.cpp with Vulkan, which supports graphics cards from NVIDIA, AMD and Intel.
What happens if a model does not fit?
Then all or part of it runs on the processor and system RAM. That works, but it is slower – how much depends on the model. The small models Qwen2.5-Coder 1.5B and 3B are quite usable on the processor as well.
Why does more context need more memory?
For every token in the context the model keeps intermediate values, the so-called KV cache. For Qwen2.5-Coder 14B that is 192 KB per token – so about 6 GB on top of the model file at 32,000 tokens.
What about laptop graphics cards?
Laptop versions often have less video memory than the desktop card with the same name. In that case choose “Other – choose video memory yourself” above and enter the value from Task Manager.
VibeCoder Studio helps you choose a model
On first start, VibeCoder Studio checks your PC and suggests models that suit your hardware. Free plan to try it out.