Zareef.com

AILane: Which Local AI Models Can Your Hardware Run?

ailane (AI Lane) is my free, open-source command line tool that inspects a machine’s hardware and tells you which local AI models it can run. It answers the questions everyone has before running AI models locally: what CPU, RAM, and GPU does this machine have? Can it run llama3.1:8b, and at which quantization? Which models should I run on it? Is my tooling (Ollama, CUDA, PyTorch) set up correctly?

It works on Windows, Linux (including WSL2 with GPU-passthrough detection), and macOS (including Apple Silicon unified memory). Zero configuration, no install required:

npx ailane systeminfo

Or install it globally:

npm install -g ailane

What it does

  • ailane systeminfo: CPU, memory, GPU, OS, and disk overview.
  • ailane gpu: deep GPU detail: vendor, VRAM, driver, CUDA/ROCm/Metal availability, WSL passthrough, Apple unified memory.
  • ailane recommend: evaluates a catalog of popular open-weight models (Llama 3.x, Qwen, Mistral, Mixtral, Phi, Gemma, DeepSeek-R1 distills) against your hardware and tiers them into best picks, runs on GPU, runs on CPU, and won’t fit, choosing the highest-quality quantization that fits.
  • ailane check <model>: can this machine run a specific model? Script-friendly exit codes, and when a model does not fit it suggests the best quantization that does.
  • ailane doctor: checks your local AI tooling (Ollama, Python, CUDA, PyTorch, Docker, llama.cpp) with install suggestions for anything missing.

A sample check:

$ ailane check llama3.1:8b

llama3.1:8b (8.03B params, q4_k_m, 8192 ctx)
  Memory needed      6.3 GiB
  Budget             11.0 GiB - NVIDIA GeForce RTX 3060 (12 GiB VRAM)

  Fits - runs on GPU with 4.7 GiB headroom.

Every command supports --json for machine-readable output, so you can gate scripts and CI jobs on whether a model fits, and --context to budget KV-cache memory for longer contexts.

How the estimates work

Memory requirements are estimated from parameter count and quantization bits per weight (measured from llama.cpp GGUF files), plus a KV-cache heuristic and runtime overhead. Budgets are deliberately conservative: 92% of VRAM, 75% of Apple Silicon unified memory, 80% of system RAM. Treat results as a good first approximation, not a guarantee.

Get it

Why I built it

Hardware sizing is the first question every private AI deployment faces, whether that is one developer’s laptop or a company GPU server. I built ailane because I kept answering this question by hand, for myself and for clients. If you are choosing between running models locally and calling cloud APIs, start with On-Premise AI vs Cloud AI, and use ailane to find out what your existing machines can already do.