Quantum math, now for AI training.
Physicists use it to fit atoms into memory. The same GPU holds a bigger training run.
Standard GPU
Your training run doesn't fit the GPU.
With Akavion The same GPU holds a bigger training run.
How it works
We store the training run more compactly.
We shrink the weights and the optimizer state. Most weights drop to 4 bits, and a correction built from quantum math reduces the quality trade-off. The optimizer keeps each direction exact and stores the sizes as a tensor train. Your data, objective, and evaluation stay as they are.
See the full walkthroughWhy it's getting worse
GPUs have plenty of computing power. They run out of memory first.
Buying more hardware helps until the next model arrives. The question that lasts is how much training you can extract per byte.
What we measure
Three numbers before we call it a result.
Memory saved, throughput cost and quality cost, measured on the same run against 16-bit and the simplest free alternative. Compression comes at a cost. We say what it is.
The trade
Choose your trade-off.
Our methods are designed to keep the cost low, and to let you pick the point on the curve that makes sense for your team.
Thesis
It starts with the math
How mathematics from quantum physics applies to AI training today.
Read the long versionEveryone else sells you cheaper chips. We make each chip do more.
Tell us what you trainWhere we are
We're early.
Akavion started with the research. We're running training jobs, measuring, and talking with teams who train models about what they design around. We're now taking on our first pilots.
Don't take our word for it.
Researchers and operators.
Two PhDs and 14 years of doctoral research between us. One of us is a named inventor on 4 patents. We've studied, published, and worked at the places below.
