Etoolio

Local AI Model Advisor

Free local AI model advisor. Auto-detects your GPU via WebGPU, then shows which local LLMs (Llama, Mistral, Qwen, Gemma) your machine can run, at which quantization, and roughly how fast. In your browser.

Checking your hardware…
Apple Silicon: use your total unified memory. No dedicated GPU? set this low (2-4).
Your machine can comfortably run models up to Gemma 2 9B on the GPU at Q4_K_M.
ModelNeedsVerdictSpeed
Llama 3.2 1B 1B2.2 GBRuns on GPUFast · ~20-60 tok/s
Gemma 2 2B 2.6B3.1 GBRuns on GPUFast · ~20-60 tok/s
Llama 3.2 3B 3B3.3 GBRuns on GPUFast · ~20-60 tok/s
Phi-3.5 Mini 3.8B3.7 GBRuns on GPUFast · ~20-60 tok/s
Qwen2.5 7B 7B5.5 GBRuns on GPUFast · ~20-60 tok/s
Mistral 7B 7B5.5 GBRuns on GPUFast · ~20-60 tok/s
Llama 3.1 8B 8B6.1 GBRuns on GPUFast · ~20-60 tok/s
Gemma 2 9B 9B6.7 GBRuns on GPUFast · ~20-60 tok/s
Phi-3 Medium 14B 14B9.5 GBPartial offloadUsable · ~8-20 tok/s
Qwen2.5 14B 14B9.5 GBPartial offloadUsable · ~8-20 tok/s
Gemma 2 27B 27B16.8 GBWon't fit
Qwen2.5 32B 32B19.6 GBWon't fit
Llama 3.3 70B 70B41 GBWon't fit
Qwen2.5 72B 72B42.1 GBWon't fit

Estimates only. Memory needs use the model size at the chosen quantization plus ~1.6 GB overhead; real usage grows with context length. Speeds vary a lot by exact GPU/CPU. Great local runners: LM Studio, Ollama, Jan. WebGPU can't read exact VRAM, so double-check the auto-detected value.

What the local AI model advisor does

Wondering “can my PC run this LLM?” This tool answers it. It auto-detects your GPU through the browser's WebGPU API, takes your memory, and shows which popular local models, Llama 3, Mistral, Qwen, Gemma and more, your machine can run, at which quantization, and roughly how fast. Everything runs in your browser; nothing is uploaded.

Why check before you download

Local model files are several gigabytes each, and a model that's too big for your memory either won't load or crawls. Checking first saves long downloads and frustration, and shows you the largest model you can run smoothly.

How to use it

  • Let it auto-detect, or set your GPU memory (VRAM) and system RAM with the sliders. Apple Silicon: use your total unified memory as the GPU memory.
  • Pick a quantization (Q4_K_M is the recommended sweet spot).
  • Read the table: each model shows its memory need and whether it runs on your GPU, partially offloads, runs on CPU only, or won't fit.

Understanding the results

  • Runs on GPU — fits in video memory, smooth and fast.
  • Partial offload — some layers on GPU, some on CPU; usable but slower.
  • CPU only — fits in RAM but not VRAM; works, but slow.
  • Won't fit — too big for your memory at this quantization; try a smaller model or lower quant.

Quantization in brief

Quantization stores model weights at lower precision to save memory. Q4_K_M is about a quarter the size of full FP16 with only a small quality drop, which is why it's the default for most people. Step up to Q5, Q6 or Q8 if you have memory to spare and want a little more quality.

Good to know

  • WebGPU can't read exact VRAM, so verify the auto-detected value.
  • Memory estimates add ~1.6 GB overhead; long context windows need more.
  • Run models with LM Studio, Ollama or Jan once you know what fits.

Good to know

Local AI Model Advisor, frequently asked questions

How do I know which AI models my computer can run?

Enter (or auto-detect) your GPU memory and system RAM, and this tool shows which popular local LLMs, like Llama 3, Mistral, Qwen and Gemma, will run on your machine, at which quantization, and roughly how fast. It estimates each model's memory need at your chosen quant and compares it to your hardware.

How does it detect my hardware?

It uses the browser's WebGPU API to detect your graphics adapter, and deviceMemory for a RAM hint. WebGPU can't read exact video memory, so it makes a conservative guess, always double-check the VRAM value with the slider. Apple Silicon users should enter their total unified memory as the GPU memory.

What is quantization and which should I pick?

Quantization shrinks a model by storing its weights at lower precision, which cuts memory use with a small quality cost. Q4_K_M is the popular sweet spot, roughly half the size of Q8 and a quarter of full FP16, and is the recommended default. Higher quants (Q5, Q6, Q8) need more memory for slightly better quality.

How is memory need calculated?

Roughly: parameters × bits-per-weight ÷ 8, plus about 1.6 GB of overhead for the runtime and context. A 7B model at Q4 needs about 5.5 GB, an 8B about 6 GB, a 70B about 40 GB. Longer context windows increase this.

How do I actually run these models?

Use a local runner like LM Studio, Ollama or Jan, they download the model in the right quantization and run it on your machine. This tool tells you what will fit before you download several gigabytes.

Are the speeds exact?

No, they're rough bands. Actual tokens-per-second depend heavily on your specific GPU or CPU, memory bandwidth and context length. Treat 'Runs on GPU' as smooth, 'Partial offload' as usable, and 'CPU only' as slow but working.