Independent Researcher Trains 1.18B LLM From Scratch on a Single Consumer GPU

Independent researcher Mert Çetin of Me Force Technology has trained CetinLM Base-v1, a 1.18-billion-parameter language model, entirely from scratch using a single RTX 4070 Ti SUPER desktop GPU with 16GB of VRAM. The model has surpassed 3.80 billion training tokens, with validation loss continuing to fall steadily from 2.592976 at 3.60B tokens to 2.577079 at 3.80B tokens, showing no signs of plateauing. Rather than relying on large-scale web data dumps, Çetin built a curated proprietary dataset over a week-long pipeline, aiming for higher logical density per token. The base model, which has not yet undergone instruction fine-tuning or alignment, reportedly runs at around 48 tokens per second on a local interface and has shown coherent responses in both English and Turkish. Çetin argues the project demonstrates that foundational language model research does not require massive compute clusters or venture capital funding to produce stable, meaningful results.