The best Ollama model for your PC is simply the most capable model that fits entirely in your GPU's VRAM. If a model fits, Ollama answers quickly. If part of it spills into system RAM, it can drop from 50 tokens per second to 5.
Below are the best models for each common VRAM size (8, 12, 16 and 24GB, plus 32GB+ and Macs), with the exact ollama run command for each. Sizes are the approximate download sizes of the default Q4 versions in the Ollama library.
Quick Answer: best Ollama model by VRAM
- 8GB (RTX 3060 Ti, 4060, 5060): Qwen3 8B for general use, Qwen2.5-Coder 7B for code.
- 12GB (RTX 3060 12GB, 4070): Gemma 3 12B or Qwen3 14B.
- 16GB (RTX 4080, 5060 Ti 16GB, 5080): gpt-oss-20b, or Qwen3 14B with long context.
- 24GB (RTX 3090, 4090): Qwen3 32B, Gemma 3 27B, or Qwen3-Coder 30B for code.
- 48GB+ / big Macs: Llama 3.3 70B.
- Q4 size ≈ 0.6 GB per billion parameters (an 8B model is about 5GB).
- Leave 1–3 GB free for context (the KV cache), more for long documents.
- MoE models (e.g. Qwen3 30B-A3B, gpt-oss-20b) are fast even when slightly too big, because only a few billion parameters are active per token.
- Check with
ollama ps: it should say 100% GPU.
💻 Prefer a laptop? See our best laptops for Ollama, ranked by VRAM, tested from 8B up to 70B models, or answer 3 questions in the Laptop Finder.
- The One Rule: The Model Must Fit in VRAM
- Best Ollama Models for 8GB VRAM
- Best Ollama Models for 12GB VRAM
- Best Ollama Models for 16GB VRAM
- Best Ollama Models for 24GB VRAM
- 32GB, 48GB and Mac Unified Memory
- How to Check a Model Fits (and Fix It When It Doesn't)
- When to Upgrade Your GPU
- Frequently Asked Questions
The One Rule: The Model Must Fit in VRAM
Ollama loads as many model layers onto your GPU as fit and runs the rest on the CPU. GPU memory is roughly 10× faster than system RAM, so even 10–20% of a model on the CPU can halve your speed. That's why a smaller model that fits completely usually beats a bigger one that doesn't.
Best Ollama Models for 8GB VRAM
GPUs: RTX 3060 Ti, 3070, 4060, 4060 Ti 8GB, 5060, RX 7600, and most gaming laptops.
| Model | Command | Size (Q4) | Best for |
|---|---|---|---|
| Qwen3 8B | ollama run qwen3:8b | ~5.2 GB | Best all-rounder: chat, reasoning (thinking mode), tool use |
| Llama 3.1 8B | ollama run llama3.1:8b | ~4.9 GB | Well-supported general model, good English writing |
| Qwen2.5-Coder 7B | ollama run qwen2.5-coder:7b | ~4.7 GB | Code completion and coding questions |
| Gemma 3 4B | ollama run gemma3:4b | ~3.3 GB | Fast, also reads images; leaves room for long context |
With 8GB, stay with 7B–8B models and keep the context window modest (8K or less). A 14B model (~9GB) won't fit and will run partly on the CPU.
Best Ollama Models for 12GB VRAM
GPUs: RTX 3060 12GB, RTX 4070, 4070 Super, 5070, RX 6700 XT, RX 7700 XT.
| Model | Command | Size (Q4) | Best for |
|---|---|---|---|
| Gemma 3 12B | ollama run gemma3:12b | ~8.1 GB | Strong writing and image understanding |
| Qwen3 14B | ollama run qwen3:14b | ~9.3 GB | Best reasoning in this tier |
| Phi-4 14B | ollama run phi4:14b | ~9.1 GB | Maths and logic |
| Qwen2.5-Coder 14B | ollama run qwen2.5-coder:14b | ~9.0 GB | Better code than the 7B, still fast |
12GB is where local AI starts to feel capable. 14B models are noticeably smarter than 8B ones. Keep context to about 8K with 14B models, or use Gemma 3 12B if you need more.
Best Ollama Models for 16GB VRAM
GPUs: RTX 4060 Ti 16GB, 5060 Ti 16GB, 4070 Ti Super, 4080, 5070 Ti, 5080, RX 9070 XT, RX 7800 XT.
| Model | Command | Size (Q4) | Best for |
|---|---|---|---|
| gpt-oss 20B | ollama run gpt-oss:20b | ~14 GB | OpenAI's open-weight model: strong reasoning, fast (MoE) |
| Qwen3 14B | ollama run qwen3:14b | ~9.3 GB | Plenty of room left for 32K+ context |
| Mistral Small 3.2 24B | ollama run mistral-small3.2:24b | ~15 GB | Instruction following; a tight fit, keep context short |
| Qwen3-Coder 30B-A3B | ollama run qwen3-coder:30b | ~19 GB | Best local coder; slightly too big, but MoE keeps it usable |
16GB is the most popular upgrade tier. gpt-oss-20b and Qwen3 14B fit fully. Qwen3-Coder 30B is about 19GB, so a few GB run on the CPU. Because it's a Mixture-of-Experts model with only about 3B active parameters, it stays usable. Expect roughly half the speed of a model that fits completely.
For coding-specific picks, see best Ollama coding models and our Qwen3-Coder Ollama guide.
Best Ollama Models for 24GB VRAM
GPUs: RTX 3090, 3090 Ti, 4090, RX 7900 XTX, and RTX 5090 laptop GPUs.
| Model | Command | Size (Q4) | Best for |
|---|---|---|---|
| Qwen3 32B | ollama run qwen3:32b | ~20 GB | Best general reasoning that fits 24GB |
| Gemma 3 27B | ollama run gemma3:27b | ~17 GB | Writing, vision, more room for context |
| Qwen3-Coder 30B-A3B | ollama run qwen3-coder:30b | ~19 GB | Coding agents and repo work, very fast (MoE) |
| DeepSeek-R1 32B | ollama run deepseek-r1:32b | ~20 GB | Step-by-step reasoning (slower, thinks first) |
24GB runs the 27B–32B class fully on the GPU. For most people these are the best models they can run locally. With 32B models, context is limited to roughly 8–16K before memory runs out; Gemma 3 27B leaves more room.
32GB, 48GB and Mac Unified Memory
- 32GB (RTX 5090 desktop): the same 32B models with 32K+ context. 70B still doesn't fit.
- 48GB (RTX 6000 Ada, or two 24GB cards):
ollama run llama3.3:70b(~43GB) fits. - Macs: unified memory works like VRAM, but macOS reserves part of it, so plan on using about 65–75% for models. A 64GB Mac can run 70B; a 128GB Mac can run
gpt-oss:120b(~65GB). See the best Mac for local LLMs.
Newer models follow the same math. New families keep arriving (see our Gemma 4, Qwen 3.5 and DeepSeek V4 guides). Check the size on the model's Ollama page and apply the rules above.
How to Check a Model Fits (and Fix It When It Doesn't)
- Run the model, then in a second terminal run
ollama ps. The PROCESSOR column should read 100% GPU. Anything like "30%/70% CPU/GPU" means it spilled over. - Too big? Pick a smaller size of the same family (e.g.
qwen3:14b→qwen3:8b), or reduce the context window. - Close other GPU apps (games, browsers with hardware acceleration, video editors). On a laptop, plug in so the GPU runs at full power.
When to Upgrade Your GPU
Upgrade when the model you want doesn't fit, not when it's merely "slow". The jumps that matter:
- 8GB → 16GB: lets you run 14B models and gpt-oss-20b. The cheapest route is an RTX 5060 Ti 16GB.
- 16GB → 24–32GB: unlocks the 32B class, a big quality step for coding and reasoning. The RTX 5090 is the fastest option.
1. RTX 5060 Ti 16GB — cheapest jump from 8GB to 16GB
Price changes often — check today’s price on Amazon
🎯 Best for: Moving from 8B to 14B–20B models
16GB GDDR7 · runs gpt-oss-20b and Qwen3 14B fully on the GPU
Check Price on Amazon →
2. RTX 5090 32GB — runs the 32B class with long context
Price changes often — check today’s price on Amazon
🎯 Best for: Qwen3 32B, Qwen3-Coder 30B, Gemma 3 27B at full speed
32GB GDDR7 · 1,792 GB/s · fastest consumer GPU for LLMs
Check Price on Amazon →Compare every option in our best GPU for local LLMs guide. Want a laptop instead? See the best laptops for running AI models locally.
Frequently Asked Questions
Sources & method: Model sizes from the Ollama model library (default Q4 tags, approximate download sizes); sizing rule of about 0.6 GB per billion parameters at Q4_K_M; ollama ps behaviour from the Ollama documentation — Himansh, TheAITechPulse
Affiliate disclosure: links to Amazon are affiliate links. I earn a commission if you buy, at no extra cost to you.