How to Pick the Right Model for Your RAM and Use Case
Not sure which model fits your hardware?
Try our interactive calculator — select your GPU and RAM to get a model recommendation in seconds.
Choosing the wrong model is the most common mistake new Ollama users make. Too large and your machine runs out of RAM mid-generation — or grinds so slowly it becomes unusable. Too small and the output quality isn't good enough for your task.
The right model depends on three things: how much RAM you have, whether you have a GPU, and what you're trying to do. This guide maps all three into a concrete recommendation you can run in under two minutes.
Quick decision table
Tokens/s figures are approximate, measured with a quantized Q4_K_M model on an RTX 4060 (8 GB VRAM). CPU-only speeds are roughly 3–5× slower.
| Model | Params | Best for | Min RAM | Tokens/s |
|---|---|---|---|---|
| Llama 3.2 3B | 3B | General chat, low-end hardware | 8 GB | ~45 tok/s |
| Llama 3.1 8B | 8B | Balanced everyday use | 16 GB | ~30 tok/s |
| Mistral 7B | 7B | Fast responses, coding | 16 GB | ~35 tok/s |
| Qwen3 14B | 14B | Math, reasoning, coding | 24 GB | ~18 tok/s |
| Phi-4 14B | 14B | Reasoning on a budget | 24 GB | ~17 tok/s |
| Qwen3 32B | 32B | Best open-weight quality | 48 GB | ~8 tok/s |
By use case
Coding and development
For code generation, debugging, and explaining code, Qwen3 14B is the top choice among open-weight models. It consistently outperforms same-size alternatives on HumanEval and MBPP benchmarks, and its context window handles large files without losing track.
If you have less than 24 GB of RAM, Mistral 7B is
the fastest coding model at the 7B tier — its responses feel snappier than Llama 3.1 8B for
short completions. Pull it with ollama run mistral.
General chat and writing
Llama 3.1 8B is the sweet spot for everyday conversation, drafting, and summarization. Meta's instruction-tuning gives it a natural, helpful tone that most users prefer over the more terse outputs from smaller models. It fits comfortably in 16 GB of RAM with the default Q4_K_M quantization.
If quality matters more than speed (and you have 24 GB+), stepping up to Phi-4 14B gives noticeably better long-document reasoning without the full cost of Qwen3 14B.
Low-end hardware (8 GB RAM or less)
Llama 3.2 3B is the model for machines with 8 GB of total RAM (or 8 GB VRAM on a dedicated GPU). At Q4_K_M quantization it uses under 2.5 GB, leaving room for your OS and browser.
On pure CPU with no GPU, expect 5–10 tok/s on a modern Ryzen or Core processor —
slow but usable for non-interactive tasks like summarization. For even smaller footprint,
ollama run llama3.2:1b
uses under 1 GB and still produces coherent text for simple Q&A.
Understanding quantization
Quantization reduces the number of bits used to store each model weight, trading a small amount of quality for a large reduction in RAM and disk usage. Ollama supports several levels:
| Level | Bits/weight | Size (14B model) | Quality loss |
|---|---|---|---|
| Q4_K_M | 4-bit | ~9 GB | Minimal — recommended |
| Q5_K_M | 5-bit | ~11 GB | Very low |
| Q8_0 | 8-bit | ~15 GB | Near-lossless |
| FP16 | 16-bit | ~28 GB | None (original) |
To pull a specific quantization, append it after the model tag:
# Q4_K_M — best size-to-quality ratio (recommended for most users)
ollama run qwen3:14b-q4_K_M
# Q8_0 — higher output quality, uses ~15 GB RAM
ollama run qwen3:14b-q8_0ollama ps.
It shows the model name, size loaded into memory, which processor (CPU/GPU) is handling it,
and how long until Ollama unloads it automatically.
# Check which models are loaded and how much RAM they're using
ollama ps
NAME ID SIZE PROCESSOR UNTIL
qwen3:14b-q4_K_M a8fbc91d2e3a 9.5 GB 100% GPU 4 minutes from nowWill your hardware handle it?
Now that you've picked a model, check whether your GPU or CPU is fast enough to run it at a useful speed. Our Hardware guide benchmarks the most common setups so you know what to expect before you download 15 GB.
Go to Hardware guide →