whichLlmmodel
Hardware Intelligence

Find local models that fit your hardware

We do the math of checking if a model will run on your hardware, so that you don't have to. Zero guesswork, zero crashed systems, zero wasted 20GB downloads.

Copy this into your coding agent
Help me find and set up the best local AI model for my hardware: fetch https://whichllmmodel.com/agent-setup.md and follow it.
Your AI agent will automatically fetch verified VRAM requirements & Ollama commands without hallucinating.
The Problem

Running local AI is full of guesswork

Between context windows, weights, and quantization levels, knowing what actually runs on your PC shouldn't require trial and error.

01 / EXPLOSION & OOM

Local models are exploding

There are hundreds of open-source model variants releasing every week. But users don't know which ones will actually run smoothly on their machine without throwing fatal Out-Of-Memory (OOM) errors.

02 / COMPLEX MATH

Calculating the math is hard & tedious

Accurately checking fit across different model architectures (GQA, MLA, Sliding Window), quantizations (Q4 vs Q8 vs FP16), and context lengths requires tedious memory equations that most developers don't have time to compute.

03 / TRIAL & ERROR

Downloading the wrong model wastes hours

Downloading a 20GB to 40GB model only to find it freezes your system or crawls at 1 token/sec is frustrating. Worse, you might miss out on a great model that your GPU could have run effortlessly.

The Solution

Precision hardware intelligence for local AI

Stop guessing. Our engine analyzes the internal architecture of every model to give you deterministic, guaranteed compatibility answers.

01

Multi-model fit check in seconds

Instantly scan our complete catalog of open-source models against your hardware. Know in milliseconds which models run at pure VRAM speeds, CPU offloading, or will OOM.

Real-time quantization matrix (Q4_K_M, Q8, FP16)
02

Architecture-aware KV cache engine

Most calculators ignore KV cache. We calculate exact KV cache memory based on your target context window (4k to 128k+) and the model's actual attention mechanism.

GQA, MLA, Sliding Window, & Mamba state sizing
03

True 3-part VRAM memory modeling

We compute true total consumption: Model Weights + Context KV Cache + CUDA/Runtime Overhead Buffers. No unexpected crashes when your prompt grows.

Weights + KV Cache + Engine Overhead = Zero OOM
Technical Credibility

Why trust our calculations?

We don't use simplified napkin math. We model the full physical memory footprint of real model runs.

No generic rules-of-thumb

We don't rely on blanket napkin formulas like “params × bytes”. Modern models frequently use mixed precision and don't always ship in standard FP16, making it impossible to accurately predict their size using generic math.

Per-architecture KV cache algorithms

Different model families use distinct attention mechanisms (GQA, MLA, Sliding Window, SSM/Mamba) which dictate how KV cache expands. We implement dedicated algorithms for each architecture to calculate exact KV cache growth.

Calibrated on real engine runs

Every formula is stress-tested against real hardware runs to guarantee precision.

How We Model Memory: Real Example
Llama 3.1 8B (Q4_K_M)
1. Model Weights (Q4_K_M)Base parameters stored in memory
4.9 GB
2. Dynamic KV CacheCalculated using GQA 8 KV heads & context length
+ 4.0 GB
3. CUDA & Engine OverheadRuntime memory workspace buffers
+ 0.7 GB
Total VRAM Needed: 9.6 GBFits safely in 12GB+ GPU (e.g. RTX 3060 / 4070)
Fits in VRAM
Napkin math (4.9 GB) misses the 4.0 GB KV cache + 0.7 GB overhead, turning a 9.6 GB load into an OOM crash.
Interactive Simulator

Test your hardware compatibility live

Moving sliders updates memory math instantly

Hardware & Context Setup

GPU VRAM
16 GB
System RAM
32 GB
Target Context Window
8k tokens

Instant Architecture Evaluation

Meta AIQ4_K_M

Llama 3.1 8B

Total: 6.3 GB (Weights: 4.8 GB + KV: 1.0 GB)
100% VRAM (Safe Fit)
AlibabaQ4_K_M

Qwen 2.5 14B

Total: 11.1 GB (Weights: 8.8 GB + KV: 1.5 GB)
100% VRAM (Safe Fit)
DeepSeekQ4_K_M

DeepSeek R1 Distill 32B

Total: 23.4 GB (Weights: 19.7 GB + KV: 2.0 GB)
CPU Offload (Slower)
Meta AIQ4_K_M

Llama 3.3 70B

Total: 48.4 GB (Weights: 42.4 GB + KV: 2.5 GB)
Out of Memory (OOM)
Want to test other quantizations (Q8, Q2, FP16) or look up a specific model?

Stop guessing. Run the right model.

Whether you want to find what fits your GPU or verify a specific model, we make it instant and deterministic.

Hardware → Models

Find models for your machine

Input your graphics card VRAM and system RAM to see every quantized model that fits at full speed.

Open Local Finder
Model → Hardware Check

Check a specific model

Look up latest open models to see exact memory math and hardware compatibility.

Search & Check Models