Same answers, smaller card
The memory moves off the GPU into RAM, so a job too big for the card still runs — and the answers come out identical. Only the memory moves; the intelligence stays put.
Nadi
Run long-context AI on a GPU that's too small for it. Nadi parks the model's memory in ordinary RAM, so the job fits — same answers as before.
Big jobs need big GPUs. Nadi moves the model's working memory off the GPU into normal RAM — so a long context that would never fit now runs on the card you already have.
| What Nadi does | What it means for you |
|---|---|
| Fits a big job on a smaller GPU | Run long-context work the card couldn't hold before |
| Keeps the model's answers | Identical to the original — nothing gets dumber |
| Costs some speed, on your terms | Slower only while the memory sits in RAM; a switch you control |
Early results on real open-source models — the win is memory, the cost is speed. Not a product.
How much GPU memory the offload saves, and whether it stays fast enough to be worth it.
When it becomes real
Nothing here is a product. No download, price, or date — a release exists only once there's a build you can run.
The usual way to run a bigger job is a bigger GPU. Nadi keeps your card and moves the memory instead.