Supra 2
Supra2-100M-Instruct · our flagship model.
Params100M
Context2,048
LicenseApache-2.0
Supra2-100M-Instruct is a 100M-parameter decoder-only model pretrained from scratch on 30B tokens of English web text, then full-finetuned for chat on a single RTX 5060 Ti 16GB. Weights are open, and the hosted demo runs this same checkpoint.
| Specification | Value |
|---|---|
| Architecture | Qwen3, decoder-only |
| Parameters | 100,684,032 |
| Layers | 12 |
| Hidden size | 768 |
| Attention heads | 12 (4 key/value) |
| Vocabulary | 32,768 |
| Context window | 2,048 tokens |
| Pretraining | 30B tokens, English web text |
| Chat format | ChatML |
| Formats | safetensors · GGUF F16 |