Research / Model releases
Models
Released base-model experiments, with current research kept separate from completed artifacts.
Released / Experimental151M
Alethic-151M
Pretraining and SinGatedAttention exploration. Original context: 256 tokens. SentencePiece Unigram tokenizer, approximately 20K vocabulary.
Historical release116M
116M predecessor
Decoder-only pretraining from scratch. Approximately 115,944,960 parameters. Released under the historical Ruzz repository name.
Hugging FaceAlethic Phase I
Approximately 1B parameters and 20B training tokens are targets. There is no Phase-I release link until a public artifact is supplied.
Current research overview