analyzeronline
Star on GitHub

$ ./can-i-run --model <hf-id>

Can I runon my machine?

Paste any Hugging Face model. Get an instant verdict, RAM/VRAM math, the best run method, and copy-paste commands for Ollama, llama.cpp, vLLM & Transformers.

tap a model to analyze it instantly

YOUR HARDWARE

Pick your parts, choose a preset, or auto-detect.

GB
GB

What can your rig run?

Popular Hugging Face models, ranked by downloads and scored against your hardware.

WHAT RUNS ON WHATrough planning guide

Laptop · no GPU
8–16 GB RAM
≤ 3B comfortably (Q4, CPU)
RTX 3060 / 4060
12 / 8 GB
7–8B (Q4) · 3B (FP16)
RTX 4070 / 4080
12–16 GB
13B (Q4) · 8B (Q8)
RTX 4090 / A5000
24 GB
32B (Q4) · 13B (FP16)
Apple M-series
32–128 GB unified
up to 70B (Q4)
Data-center GPU
A100 / H100 80 GB
70B (FP16) · 400B+ (multi)

QUANT CHEAT-SHEET

Memory ≈ params × bytes/param. Lower bits = less VRAM, slightly less quality.

FP16 / BF162.0 B/paramfull quality
INT8 / Q81.0 B/param~lossless
Q5_K_M0.7 B/paramgreat
Q4_K_M0.55 B/paramthe sweet spot
Q3_K0.42 B/paramnoticeable loss
Q2_K0.33 B/paramlast resort
01
Paste a model
Any Hugging Face URL or owner/model ID.
02
Set your rig
Type specs, pick a preset, or auto-detect.
03
Get the verdict
Memory math, run method, copy-paste commands.

Local AI model hardware requirements

Wondering how much VRAM a model needs, or whether your GPU can run it? Every popular model and GPU below has a dedicated page with exact memory requirements at each quantization, per-GPU verdicts, and copy-paste commands for Ollama, llama.cpp and vLLM. New here? Start with the LLM VRAM requirements guide.

Popular models — can you run them?

What can your GPU run?

Running AI models locally — FAQ

How do I know if I can run an AI model locally?

You need enough GPU VRAM (or system RAM for CPU / offload) to hold the model weights plus overhead. A rough rule: at 4-bit quantization a model needs about 0.6 GB per billion parameters, so a 7B model needs ~5 GB and a 70B model ~40 GB. Paste a Hugging Face model above to get an exact verdict for your hardware.

How much VRAM do I need for a 7B, 13B, or 70B model?

At 4-bit quantization, roughly: 7B ≈ 5 GB, 13B ≈ 9 GB, 34B ≈ 22 GB, 70B ≈ 40 GB of VRAM. Full FP16 precision needs about four times as much. If you don't have enough VRAM, you can split the model across GPU + system RAM (offload) at the cost of speed.

What's the easiest way to run an LLM locally?

Ollama is the simplest — install it and run a single command like `ollama run llama3.1:8b`. llama.cpp gives more control with excellent CPU and GPU-offload support, and vLLM is best for fast GPU serving. Each model page here shows the exact command for that model.

Does quantization reduce quality?

A little. 8-bit is nearly lossless; 4-bit (Q4_K_M) is the popular sweet spot with only a small quality drop and about a quarter of the memory. Below 3-bit the loss becomes noticeable. Quantization is what makes large models runnable on consumer GPUs.

Can I run AI models without a GPU?

Yes — smaller and quantized models run on CPU using system RAM (best via a GGUF build with llama.cpp or Ollama), just more slowly. A dedicated GPU is the biggest speedup for text generation.