Open weights are not enough.
Build your own model.
Train your own ML Model using your own data and benchmarks. Join the community actively looking for the cheaper and memory-enabled successor to the transformer.
2 training runs per month · unlimited inference · cancel anytime
The distinction
Open weights are not open source.
The open-weight models — Llama, Qwen, DeepSeek, Mistral — are genuinely good. You can download them, run them locally, and fine-tune them. That's real progress, and we celebrate it.
But free weights give you inference, not agency. You can't retrain the model from scratch. You can't change the architecture. You can't swap the attention mechanism for something with memory. You're renting a pre-built brain and customizing the furniture.
Open source means you control the full pipeline:
- →The architecture — not just the weights, the design of the model itself
- →The training pipeline — how it learns, what it learns on, how it improves
- →The data — your corpus, your domain, not a generic internet scrape
- →The weights — yours to keep, version, and deploy anywhere
That's what Piro gives you. Not a model to download — a model to build.
What Piro is
The platform to train your own model.
Piro trains a tiny model — ~10M parameters — on your data, against your benchmarks. Not a fine-tune of someone else's model. A model trained from scratch, on your terms, that you own completely.
The architecture is a Continuous Thought Machine — a memory-enabled successor to the transformer. Instead of processing tokens in a single pass, it thinks in bursts: refining, recalling, and adjusting across multiple cycles. It's the research frontier for models that reason, not just predict.
You keep the weights. Query them forever. No per-token tax, no API dependency, no one can take it away.
Connect your data
GitHub repos, Gmail, Notion, anything text. Piro trains on the corpus you choose.
Train your model
We run the GPUs. You keep the weights — checkpointed and versioned.
Query it forever
Unlimited inference. Your model, your API, your cost — fixed regardless of usage.
Pro plan
One plan. One price. No surprises.
Less than one week of a frontier API habit.
- 🏋️
2 training runs / month
Full GPU training on your data
- ⚡️
Unlimited inference
Query your trained model anytime
- 🤖
Architecture Copilot (GLM 5.2)
Design help for your model classes
- 📊
Benchmark suite
Sanity, generalization, adaptive compute
- 📦
Model versioning
Roll back to any prior checkpoint
- 🔑
API access
Use your model in your own apps
Cancel anytime from your billing portal. No commitment. No per-token tax.
Stop renting intelligence.
Build your own.
Join the community training tiny, owned models on their own data — and exploring what comes after the transformer.
Start training your model →