MIRA
Join the waitlist

A workspace for building small language models

Build your own
intelligence.

Draw the architecture on a canvas, train it on your own machine, evaluate it, and export a bundle that runs without MIRA. Every number labelled calculated, measured, or estimated. Nothing leaves your computer.

See how MIRA works
decoder-study/Architecturerev 4
Selected

Transformer

Each block combines attention, a feed-forward network, and normalization. Add layers and the parameter cost updates.

Layers
6
Attention heads
6
Hidden size
384
Context
1,024
Total parameters17.33Mcalculated
Interactive: click a block, add a layerIllustrative project · counts calculated · planned tagged

From one model
to a system you own.

01 Build

Start with a model you can see.

Lay out the architecture as blocks, not boilerplate. Open any block. Adjust layers, heads, and context, and watch parameter counts, memory, and training time update as you decide. Then press Train: MIRA writes the trainer as a readable train.py and runs it locally.

transformer block × 6
Attention6 heads · 384 wide
Add & normresidual · layer norm
Feed-forward4× width · 1,536
Add & normresidual · layer norm
Every block is the same four steps. Change one number and the parameter count follows.

02 Compose

A model is one part. See the whole system.

Data, model, training, evaluation, and deploy sit on one canvas, with every version kept. Memory and agents are the next nodes on the roadmap, marked as planned until they ship.

decoder-study/Systemrev 4
Selected

Model

The architecture you designed on the canvas. One node in a larger system.

decoder-study
6 layers · 17.33M calculated
versions
rev 1 → rev 4
Interactive: click any nodeIllustrative project · counts calculated · planned tagged

03 Own

Everything you build stays yours.

Architecture, trainer, tokenizer, weights, evaluations: one folder of plain files on your disk. The exported bundle runs without MIRA and refuses a drifted tokenizer. No accounts, no cloud, no lock-in.

Today, scattered
notebook.ipynbcell 47 · model = build(cfg) # which cfg?
config.yamllayers: 6 # edited by hand, unversioned
terminalrun_47.log · loss numbers · no trainer kept
In MIRA, one exportable folder
decoder-study/rev 4 Export
architecture.json
decoder · 6 layers · 17.33M params, calculated
train.py
the trainer MIRA generated · readable · rerunnable
tokenizer.json
BPE · fingerprint checked on load
data/
your corpus · train / eval split
weights/
checkpoints per version · rev 1 → rev 4
evals/
12 cases · results per version
generate.py
runs the model without MIRA
Plain filesRerunnableYours

— The MIRA system

One workspace.
The whole small-model stack.

Data flows into a model you designed. Training runs locally. Evaluation runs real forward passes. Deploy exports a bundle that stands alone. Recipes, experiments, and versions connect every stage. Memory and agents come next, and are marked planned until they do.

  1. DataYour files, split and tokenized on your machine.
  2. ModelsArchitectures you lay out block by block.
  3. TrainingA readable train.py, run on your own hardware.
  4. EvaluationReal forward passes against a case suite.
  5. DeployA bundle that runs without MIRA.
  6. MemoryplannedStores the model and agents read and write.
  7. AgentsplannedA trained model plus tools and a goal.
  • RecipesStarter configurations you can import and edit.
  • ExperimentsSweeps as real runs, each priced before launch.
  • VersionsEvery snapshot of the model, with restore.
  • LibraryplannedEvery model and dataset with where it came from.

For people who build
the model, not just call it.

ML researchers who want the architecture in view. Founders training small models on their own hardware. Teams who would rather own the weights than rent an API.

Build the intelligence you’ll own.

MIRA opens in groups. Join the waitlist and we’ll send details when yours does. Runs on your machine; nothing leaves it.