Simplifying model finding for your specific hardware
Our Mission
The open-source AI ecosystem is exploding. Every single week, dozens of new foundation models (DeepSeek, Llama, Qwen, Gemma, Mistral) and quantization formats (Q4_K_M, Q8_0, FP8) are released.
Yet for developers and enthusiasts, this explosion creates immense friction: Which of these models can my hardware actually run? Spending 45 minutes downloading a 25GB GGUF model only to have your GPU throw a CUDA Out of Memory (OOM) error — or crawl at 1 token/sec — turns what should be an exciting experience into frustrating guesswork.
Most online calculators use naive formulas that ignore how modern attention architectures actually allocate memory, failing on dynamic KV caches and partial CPU offloads.
We built whichllmmodel to give developers instant, empirical clarity. By calculating exact layer weights directly from model byte files, modeling dynamic KV caches across 8 attention shapes, and computing real GPU/RAM offload splits, we tell you exactly which models will run at peak speed on your hardware before you ever hit download.
Zubair Tahir
Founder & Lead Developer
Hey, I'm Zubair! I'm a software engineer who loves building things that make human lives easier. When trying to run models locally on my own hardware, I got tired of guessing quantization sizes, doing manual KV cache math, and dealing with frozen systems from memory overflows. I wished there was a tool where I could just enter my GPU VRAM and System RAM, set my context length, and immediately see which models run at full speed versus CPU offload. I built whichllmmodel to solve that problem for myself and the developer community. Now, it has made my life so much easier and faster for selecting and using local models. If it saves you from a single OOM crash, that makes my day!