The second rung of the ladder: 178 million parameters, warm-started from G-Micro, trained in one week of free GPU time.
Not public yetSource code
| Status | Building |
|---|---|
| Kind | AI model |
| Year | 2026 |
| Built with | PyTorch, Python |
The successor to G-Micro, aimed squarely at the things G-Micro gets wrong, with a hard budget: a single week of free Kaggle quota on two T4s. Every choice in the configuration came from a measurement rather than a guess.
Warm-starting from the smaller model rather than beginning at random is what makes the budget work, the new network inherits what the old one already learned about Polish and spends its week on what it does not know.
loading...
AI guide. It can make mistakes.