50GRAMx Think

Think.

Runs open AI models (text, speech, vision) on your grams, with every answer signed by the machine that made it.

  • Every step signed where it ran
  • Measured on real machines first
  • Your data stays on your grams
  • Energy measured, not modelled
Does the job ofOpenAI APIAmazon BedrockTogether AIReplicate
A signed measurement of this machine
$ epnd bench cpu
{
  "result": {
    "metric": "cpu.throughput",
    "value": 153.80340074778266,
    "unit": "hashes/s",
    "node_did": "did:epn:0024080112208414…bd024bfdd68672",
    "extra": { "cpu_model": "Intel(R) Core(TM) i5-5350U CPU @ 1.80GHz", "cores_physical": 2 }
  },
  "signature": "A+lGCMXcikiDAoscA+e3wN8xyuPsl8Mn+ndf982Po8jf0FLa…"
}
cpu: 154 hashes/s over 507ms
Real output. Run on the founder’s Mac gram, 16 Nov 2026. The gram’s id and signature shortened.
In production
719signed runs on the network50gramx.com, 16 Nov 2026
152models measuredsame reading
35machines that signedsame reading
21model providers coveredsame reading
See it for yourself →

Everything Think does

All features

What you get

  • ✓AI on hardware you own, with the speed of each model measured on real machines first.
  • ✓Private data never leaves your grams.

Why it is different

  • →Every answer is signed, so you can show which machine and which model made it.
  • →The network measured these models before you choose one.

Made forai buildersdeveloperssmall businesses

How it works

The whole flow
1You ask2A gram is chosen3The model runs4The step is signed5You get both

Start in a few commands

  1. 1. Measure this machine
    $ epnd bench inference
  2. 2. Ask Evo
  3. 3. Check a result
    $ epnd bench verify result.json
Read the quickstart
A signed measurement of this machine
$ epnd bench cpu
{
  "result": {
    "metric": "cpu.throughput",
    "value": 153.80340074778266,
    "unit": "hashes/s",
    "node_did": "did:epn:0024080112208414…bd024bfdd68672",
    "extra": { "cpu_model": "Intel(R) Core(TM) i5-5350U CPU @ 1.80GHz", "cores_physical": 2 }
  },
  "signature": "A+lGCMXcikiDAoscA+e3wN8xyuPsl8Mn+ndf982Po8jf0FLa…"
}
cpu: 154 hashes/s over 507ms
Real output. Run on the founder’s Mac gram, 16 Nov 2026. The gram’s id and signature shortened.

Questions people ask

All questions
Which models can I run?

Open models the engines support: text models through llama.cpp, Ollama and vLLM, plus speech engines. The inference pages list every model measured on the network.

Can I call it from my own app like the OpenAI API?

Inside a gram, the engines speak the OpenAI API’s shape. An endpoint you call from outside your grams is not open yet.

Does my data leave my machines?

No. The model runs on your grams, where your data already is.

Start with Evo

Evo works on your own gram, with your machines in view, and asks before every change. Take one of these to it, or write your own.

  1. 1. Run a gramOn a machine you have. Evo lives there.
  2. 2. Enter your gramxYour business on the network.
  3. 3. Ask EvoPaste the question; it works through it with you.
Start with ThinkThe OpenAI API shape the engines already speak inside a gram.
Works with