Supra 2
Supra2-100M
The one to start with. A 100M decoder-only model, pretrained from scratch on 30B tokens of English web text with a custom 32,768-token tokenizer. Base is the raw model. Instruct is a full finetune of that base, done on a single RTX 5060 Ti 16GB.
Supra2-IMG
A 104M diffusion transformer, trained from scratch. Frozen Flan-T5-Base, SD-VAE, 256² images. Ten epochs over 5.6M synthetic images, on one H100, in about nine hours. State of the art for the size — and still small enough to hold.
Supra2-Medium
25 million parameters, trained on 20B tokens — about 800 tokens per parameter. A 16,384-token tokenizer and a 1K context. Small enough to study, large enough to speak in short turns. Base for continuation. Instruct for chat.
Supra2-Nano
An 800K-parameter Qwen3 model trained on 1B tokens, with a 4,096-token vocabulary so the embeddings do not eat the budget. A baseline for the family, not a chatbot. Coherent long text is not the point. The pipeline is.