RAMDeck research
Answer me with artifact · 2026-10-28 08:07

New version 0.5.10

You have 0.5.5. Found when this page was made.

What's new:

Run in a terminal:

npx skills update answer-me-with-html -y

Or tell your agent:

Update answer-me-with-html
See what's new

RAMDeck research

Lead: RAMDeck is a $109 Indiegogo hub from TradeMAV. It splits one model's layers across the PCs, Macs and phones on your LAN. The engine underneath is llama.cpp RPC, which is free. What you pay for is the setup layer: auto discovery, layer placement, a dashboard and installers. It lets a model run that fits on no single machine. It does not make that model fast.

A

What it is

Product
RAMDeck Hardware Hub + node apps (Win, macOS, Linux, iOS, Android)
Maker
TradeMAV (a small trading-tools company)
Price
from $109 on Indiegogo; campaign ends 2026-11-08
Hub hardware
specs not disclosed
Engine
llama.cpp ggml-rpc-server (ports 50052 GPU, 50053 CPU)
API
OpenAI-compatible endpoint on the hub
Public code
ramdeck-core-public, v0.9, Python, 67 stars. Node agent only; the coordinator is proprietary
License
Apache 2.0 + Commons Clause, so no commercial use or resale
B

How it works

layers 1-20layers 21-40layers 41-60activations over LANactivations over LANnext tokenHubMacPC with RTX 3060Old laptop

Each token goes through every device in turn. The network and the slowest node set the speed. The memory is not one shared pool.

C

Claimed numbers

Model Devices Speed
13B Qwen3.5 laptop + Mac + RTX 3060 11-12 t/s
40B 3 16 t/s
120B 5 (~111 GB) 3.1 t/s
120B 6 1.37 t/s

These numbers come from the vendor and the tweet. Nobody has tested them independently. The 120B results show speed going down as devices are added. Reddit users report the same thing with llama.cpp RPC: adding a second host halved MoE speed.

D

Verdict for Allternit

Watch, do not buy or integrate
  • The free tools do the same job: llama.cpp --rpc, exo and distributed-llama.
  • Commons Clause blocks it from any paid Allternit product.
  • llama.cpp calls its RPC backend a proof of concept and says never to expose it on an open network.
  • Useful idea: Allternit already has paired computers on the mesh (Test MacBook, Mail VPS, Eoj's cloud node). llama.cpp RPC across them could run an open model locally for dogfooding, at low t/s.
  • It does not fit the cloud inference plan. Production serving needs one GPU node (SGLang/DeepInfer), not a LAN chain.
What next for this link?