INDEPENDENT AI RESEARCH

Small-scale research.
Published honestly.

Sinople is an independent AI research project building models from scratch and publishing every experiment — including the runs that fail.

CURRENT RELEASE

Viperthon Gen-1 B′

A small text-continuation model trained entirely from scratch.

Parameters 31.5M trained from scratch
Training budget 344M / 415M tokens, training in progress
Held-out accuracy ~48.9% token accuracy
Training data smoltalk no pretrained backbone

ABOUT THE MODEL

Viperthon is a text continuation model.

Viperthon Gen-1 B′ generates English prose from text context. It is not an assistant and is not designed to answer questions, follow instructions, or remember a conversation.

The model was trained from scratch with no pretrained backbones and no fine-tuning. Its weights are publicly released.

RESEARCH PRINCIPLES

Every run is part of the record.

01

From scratch

Models are designed and trained from scratch rather than built on pretrained backbones.

02

Full experiments

Failed runs are published alongside successful ones, keeping the experimental record intact.

03

Reproducible

Experiments are designed to be reproducible on free-tier compute.

VIPERTHON GEN-1 B′

Read the record or run the demonstration.