Don’t Believe Us. Prove It Yourself.
Every RAMDeck cluster ships with its own built-in benchmark button. But while you wait, here are some raw, unedited proofs of our own machines.
1. Running a 13B AI Model Across 3 Home Devices
We split a 13B parameter AI model across three devices already on hand — an old laptop, a mini PC with an RTX 3060, and a Mac mini — running at 11–14 tokens/sec over plain Ethernet.
2. Loading a Massive 40B Model Across 3 Devices
Pushing further: we load a massive 40B parameter model (~25GB) across a 3-device cluster. Full load completes in roughly 2.5 minutes, running at approximately 16 tokens/sec.
3. Loading a 27B Model on an Old Laptop
A 27B parameter model is roughly 16GB on its own — more than the 12GB Windows laptop used as the primary node could hold alone. RAMDeck splits it across four devices with the laptop staying primary.
4. Adding an Android Phone to a Live Cluster
We add an Android phone as a genuine fourth contributing node while the cluster is live, showing the automatic rebalance triggered by the new device joining.
5. Coding from Another Room
We loaded a 7B model from the 12GB Acer laptop. RAMDeck routed the inference to the RTX 3060 PC, hitting 43 tok/s. We coded from a Mac Mini — a third machine — using Continue.dev pointed at RAMDeck’s API, successfully summarizing a multi-thousand-line file.
6. Indexing a RAG Knowledge Base
We proved it on the same 12GB Acer laptop holding the durable backup index. Before ingestion, the chatbot invented a wrong answer about our own product; one FAQ file later, it answered correctly, straight from the document.
Take Back Control of Your AI
Our pre-launch campaign is now live. Join the waitlist on our Indiegogo page to get priority access when we launch.
View Indiegogo Campaign