Model picker
Which local AI model fits your computer
Local AI models range from 500 MB to hundreds of gigabytes, depending on size. A model too heavy for your hardware will hang or fail to load at all. Pick your GPU from the list or drag the slider — the calculator instantly tells you which model will actually run.
Looks like Gemma 4 E4B (~4B active (MoE)) would fit
On a CPU the model runs noticeably slower than on a GPU — fine for short everyday requests, can get annoying in long conversations.
The full model catalog
Real GGUF quant sizes (Q4_K_M) pulled from Hugging Face — not eyeballed estimates.
| Model | Parameters | File size | Memory needed | Tier |
|---|---|---|---|---|
| Qwen3.5 0.8B | 0.8B | 0.53 GB | ≥2.1 GB | very weak hardware |
| Llama 3.2 1B Instruct | 1B | 0.81 GB | ≥2.5 GB | very weak hardware |
| Qwen3.5 2B★ | 2B | 1.28 GB | ≥3 GB | light |
| Llama 3.2 3B Instruct | 3B | 2.02 GB | ≥3.9 GB | light |
| Phi-4 mini Instruct | 3.8B | 2.49 GB | ≥4.5 GB | medium |
| Qwen3.5 4B★ | 4B | 2.74 GB | ≥4.8 GB | medium |
| Gemma 4 E2B | ~2B active (MoE) | 3.11 GB | ≥5.2 GB | medium |
| Mistral 7B Instruct v0.3 | 7B | 4.37 GB | ≥6.7 GB | medium |
| Gemma 4 E4B | ~4B active (MoE) | 4.98 GB | ≥7.5 GB | medium |
| Qwen3.5 9B★ | 9B | 5.68 GB | ≥8.3 GB | powerful |
| GigaChat3.1 10B A1.8B | 10B (1.8B active, MoE) | 6.47 GB | ≥9.3 GB | powerful |
| Gemma 4 12B | 12B | 7.12 GB | ≥9.9 GB | powerful |
| Mistral Nemo 12B Instruct | 12B | 7.48 GB | ≥10.5 GB | powerful |
| Phi-4 14B | 14B | 9.05 GB | ≥12.4 GB | powerful |
| Gemma 4 26B A4B | 26B (4B active, MoE) | 16.95 GB | ≥21.8 GB | top-end |
Not included in the local picks
Even the lightest usable quant doesn't fit consumer hardware — available through Doka Cloud instead.
- DeepSeek V4 Flash — Младший вменяемый квант (IQ2_XXS, уже заметная потеря качества) — 86,7 ГБ. Более лёгкие файлы в репозитории (3,8–6 ГБ) — это MTP/draft-модули, а не полная модель.
- Llama 4 Scout 17B-16E — MoE с 16 экспертами — даже младший квант (TQ1_0, экстремальная потеря качества) весит 29,3 ГБ. Название «17B» — активные параметры, реальный вес на диске в разы больше.
Frequently asked questions
How accurate is this calculator?
We use the official VRAM size for each card model — a fixed number from the manufacturer, no guessing involved. Accuracy depends on picking the right model (and, for cards with multiple VRAM versions — like the RTX 4060 Ti at 8 or 16 GB — the right one). Every model in the catalog already has headroom built in for OS overhead and conversation context, so the recommendation isn't cutting it close.
Can I run an AI model without a graphics card?
Yes — on integrated graphics or with no GPU at all, the model runs on the CPU. It's slower, but perfectly usable for everyday tasks with lighter models (1–4B parameters). Pick “No GPU” and set your RAM with the slider.
What's MoE, and why does it still need as much memory as a full model?
MoE (mixture-of-experts) is an architecture where only part of the weights are active per token, which makes inference faster. But all the weights still have to fit in memory at once — MoE saves compute, not RAM or VRAM.
Why isn't DeepSeek or Llama 4 Scout in the calculator?
Both models weigh 30 to 90+ GB even at the lightest practically usable quant — well beyond any consumer computer's memory. You can still use them through a cloud provider in Doka, just not locally.
I have a Mac — how do I use this?
On Apple Silicon (M1–M4), memory is unified between CPU and GPU — there's no separate VRAM limit, the whole RAM pool is available to the model. Pick “No GPU” and set the slider to your Mac's total RAM (usually 8, 16, 24, or 32 GB).