Skip to content
NewRedrob ImageOpen-weight image model on Hugging Face

Open-weight models thatrun on your laptop.

Small enough for hardware you already own.

Clone the projectsAbout the lab
Grainy illustration of a laptop, a phone, and a single graphics card on a desk, with charts pinned on the wall above them

How the pieces fit

One question, four projects

Can a model small enough to run on your own hardware do the job?

  1. 01

    Decide what good enough means

    Without a bar there is no way to say which model you need. Settle it first, on your own task and in your own language.

    • Redrob EvalOut now

      Any model from any source on your prompts, settled by blind preference and relative cost.

  2. 02

    Build a model that clears it

    And do it at a size that still fits one machine.

    • Redrob ImageOut now

      Open-weight text-to-image.

    • Redrob VerifyOut now

      Open-weight OCR, forgery, and face models for identity documents.

    • Small language modelIn training

      Hindi, Korean, and English, sized to finish on one GPU.

    • Vision-language modelNext

      Reads a screen or a document on the device showing it.

  3. 03

    Run it on your own machine

    Put it where the data already is, with nothing routed through anyone else's API.

    • Redrob StudioOut now

      Image and audio tools on your own GPU.

Why it matters

Cheaper, offline, and yours

No The chunks of text an AI reads and writes. Most AI services bill you by how many chunks you use. bill

Running it costs nothing beyond the hardware you already own. Use Redrob Eval to see how much you would save.

Works with the network off

The model runs where the data already is, so a dead connection or a rate limit does not stop the work.

Nothing leaves, nothing gets revoked

An open license that lets anyone use, change, and sell what they build on it, as long as the credit notice stays. The file a model ends up as after training. Hold the file and you can run the model yourself. on your own disk. Nobody can swap the model under you or switch it off.

We are not claiming a small model beats frontier models on every task, only that it can clear the bar on yours. Use Redrob Eval to compare it on your own prompts.

For builders

One command.
Every project.

Public projects, quickstarts copied straight from the READMEs.

Open projectsAbout the lab

4

Public projects

Apache-2.0

Same license across all four

ko / en

Both locales, equally

git clone https://github.com/redrob-labs/redrob-eval.git && cd redrob-eval && yarn install && yarn dev

1 GPU

the budget we design against

10M+

people on the Redrob platform

0

signups, API keys, or permissions

Two ways to start

Build on your own

Clone it and run it.

  • Public GitHub projects
  • Apache-2.0 / MIT licenses
  • Quickstarts from real READMEs
  • Numbers with the method next to them
Start on GitHub

Talk to the lab

We wrote down what to send before you write.

  • Research collaboration
  • Joining
  • Media and talks
  • One email, no forms
Contact