A workspace for building small language models
Design, train, evaluate, and export small language models from one visual workspace. Train on your machine, or in the cloud. Every number is labeled: calculated, measured, or estimated. Nothing leaves your control.
- You own your models.
Weights, tokenizer, trainer, and evals are plain files in your folder.
- Transparent by design.
Each figure says whether it was calculated, measured, or estimated.
- Runs anywhere, no lock-in.
The exported bundle runs with Python and no MIRA install.
Embedding
- Vocabulary
- 16,384
- Dimension
- 384
- Parameters
- 6,291,456calculated
- Tensor shape
- (16384, 384)
- Memory, fp16
- 12.6 MBestimated
Illustrative configuration, not MIRA’s trained Tiny. Tiny is byte-level: vocabulary 259, parameters 11,140,224, measured
01Build
From one model
to a system you own.
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.
- Visual and modular
Design complex systems as simple blocks.
- Real numbers
See parameters, memory, and training time live.
- Yours to run
Generates readable code that runs on your machine.
Model
The architecture you designed on the canvas. One node in a larger system.
- decoder-study
- 6 layers · 17.33Mcalculated
- versions
- rev 1 → rev 4
Illustrative project: rev 4 and the version history stand in for your own. 6 layers and 17.33M parameters are the calculated default spec; 12 cases is the bundled arithmetic suite.
02Compose
A model is one part.
See the whole system.
The model you drew is one node. Planned nodes are marked.
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.
Model
The architecture you designed on the canvas. One node in a larger system.
- decoder-study
- 6 layers · 17.33Mcalculated
- versions
- rev 1 → rev 4
Illustrative project: rev 4 and the version history stand in for your own. 6 layers and 17.33M parameters are the calculated default spec; 12 cases is the bundled arithmetic suite.
03Own
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.
- Plain files
Yours to keep and modify.
- Rerunnable
Train, evaluate, resume anytime.
- No lock-in
Runs anywhere, no MIRA required.
rev 47 itemsarchitecture.jsondecoder · 6 layers · 17.33M params calculatedtrain.pythe trainer MIRA generated · readable · rerunnabletokenizer.jsonBPE · fingerprint checked on loaddata/your corpus · train / eval splitweights/checkpoints per version · rev 1 → rev 4evals/12 cases calculated · results per versiongenerate.pyruns the model without MIRA
- Plain files
- Rerunnable
- Yours
Illustrative folder: rev 4 stands in for your own version history. 6 layers and 17.33M parameters are the calculated default spec; 12 cases is the bundled arithmetic suite.
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.
Data
Your files, split and tokenized on your machine.
Models
Architectures you lay out block by block.
Training
A readable train.py, run on your own hardware.
Evaluation
Real forward passes against a case suite.
Deploy
A bundle that runs without MIRA.
MemoryPlanned
Stores what the model and agents read and write.
AgentsPlanned
A trained model plus tools and a goal.
Recipes
Starter configurations you can import and edit.
Experiments
Sweeps as real runs, each priced before launch.
Versions
Every snapshot of the model, with restore.
LibraryPlanned
Every model and dataset with where it came from.
Built for builders
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.
- Full control
Design, train, and iterate on your terms.
- Built for real work
From experiment to deployment.
- Yours to keep
You own the models, weights, and outputs.