Hardware sizing tools for production LLM deployments. Open-access engineering courses for the developers building them.

The Mission

ApX Machine Learning is built for engineers deploying open-weight models and building AI systems. Sizing hardware for production requires more than rule-of-thumb estimates that just multiply parameter counts by quantization bytes. Under actual traffic, memory and latency are governed by KV cache growth, context length, batch concurrency, and multi-GPU interconnects.

We build sizing tools that model actual production constraints and verify those predictions against physical GPU benchmarks. Alongside these tools, we maintain an open-access library of engineering courses so developers can master the math, architectures, and systems code underneath.

What We Build

Why This Exists

I originally built ApX Machine Learning as a set of AutoML tools for my own personal use. As open-weight models like Llama and Qwen became viable to run locally, I started writing guides on how to run them and what GPUs they required.

The turning point came with the launch of DeepSeek V3, when the industry realized open models could compete at the absolute top tier. At the time, ApX had the only comprehensive guide detailing the VRAM requirements and hardware constraints of running it, and traffic to the site surged.

Since then, ApX has transitioned into a dedicated engineering platform for sizing and deploying open-weight models at production scale. Today, it hosts the VRAM calculator, model directory, and reference guides used by thousands of engineers weekly.

Wei-Ming Thor

Founder

How Specifications & Guides Are Verified

Hardware specifications, memory calculations, and serving models are built from official model configuration files, architecture whitepapers, and physical GPU test runs. We continuously benchmark our calculator's memory and throughput predictions against actual hardware executions and publish our accuracy reports openly.