The question
Modern models such as Claude Sonnet 5.5, Claude Opus 5.5 and Gemini 4 Argon are strong at code and reasoning. We test whether they can create smaller siblings of themselves.
Research project by lgteamai
AI makes AI
Can Claude Sonnet 5.5 build a smaller sibling of itself?
We research whether a modern language model can build a quality language model of its own, under human supervision.
Can an AI build a good AI?
We build artificial intelligence from scratch and test how far neural networks themselves can go.
From scratch
Earlier versions were based on existing models. Now every model we make is trained from scratch.
Human supervision
A person supervises the process and checks the results, so the answer to our question stays honest.
Models
Every Nano model, from the first experiments to the current flagship.
In development
Nano-1.5-Pro-6B-A2B
Our new flagship model. A mixture-of-experts architecture with 6 billion parameters, 2 billion of them active. With it we check how well a neural network handles MoE architectures.
- Total parameters
- 6B
- Active parameters
- 2B
- Data
- Up to August 2026
- Access
- Closed
Readiness
1%
Training starts soon.
-
A 196 million parameter model trained to the Chinchilla optimum. With it we practiced following a training run together with a human. People built most of this model, and access is closed: only the team can test it.
-
An improved Nano 1.2 Lite, which came out in July. It saw far more data during training and no longer guesses at random most of the time. It scores 35.47% on ARC-Easy.
-
Released on 31 July 2026. The model was heavily undertrained and scored 26.47% on ARC-Easy, which is almost always random guessing. It is hidden from Hugging Face and fully closed, so it cannot be used. The Nano 1.2 line was abandoned because of limited compute.
-
Planned as a 1.5 billion parameter model for compact mini-coding on a phone, where it would run comfortably and handle micro tasks. Cancelled because of limited compute.
-
We fine-tuned Qwen 2.5 Coder Instruct on new data, including facts about our company and fresh information as of December 2025. The preview opened on 25 March 2026, the full release followed on 17 April 2026, and the model was shut down in mid-May 2026.
-
An early experiment that has been left behind.
-
An early experiment that has been left behind.
We have no public models to try right now. Access to every model above is closed.
Benchmarks
Results for Nano-1.2-Optimal.
| Benchmark | Nano-1.2-Optimal |
|---|---|
| ARC Easy | 38.89% |
| COPA | 63.0% |
| Release date | |
| Parameters | 196.3M |
ARC Easy across the Nano 1.2 line
Nano-1.2-Lite
26.47%
Nano-1.2-Lite-0926
35.47%
Nano-1.2-Optimal
38.89%
Scale from 0 to 50%
Random guess 25%
Like the project
0
likes in total
One like per browser. We use a cookie and a hashed form of your IP address only to prevent repeated likes.
Last likes
No likes yet. Be the first.