{"name":"ttf-home","repo":"oguzeldereli/ttf-home","url":"https://midorreal.com/project/oguzeldereli-ttf-home","repository":"https://github.com/oguzeldereli/ttf-home","description":"Ternary language models on a home computer: run, train, convert, fine-tune, export to GGUF","facts":{"language":{"value":"C++","source":"https://github.com/oguzeldereli/ttf-home","checked":"21 Nov 2026"},"last_release":{"value":"v0.1.0, 9 Nov 2026","source":"https://github.com/oguzeldereli/ttf-home/releases","checked":"21 Nov 2026"},"releases":{"value":"1","source":"https://github.com/oguzeldereli/ttf-home/releases","checked":"21 Nov 2026"},"contributors":{"value":"1","source":"https://github.com/oguzeldereli/ttf-home/graphs/contributors","checked":"21 Nov 2026"},"archived":{"value":"no","source":"https://github.com/oguzeldereli/ttf-home","checked":"21 Nov 2026"}},"stars":30.0,"writeup":{"what_it_is":"ttf-home is a command-line tool (`ttf`) for running, training, converting, fine-tuning and exporting ternary language models (weights in {-1, 0, +1}) on a home computer. It uses the TTF trainer and llama.cpp with TTF patches for inference, and exports models as GGUF files.","audience":"Users who want to run or train ternary language models on personal hardware with NVIDIA GPUs or CPUs","claims":[{"kind":"broad","claim":"Trains with weights kept in {-1, 0, +1} throughout, using a low-rank accumulator instead of master weights and Adam states, so training needs less memory","excerpt":"the weights stay in {-1, 0, +1} the\nwhole time, with a low-rank accumulator in place of master weights and Adam states, so training needs a fraction of the\nusual memory","status":"too_broad"},{"kind":"specific","claim":"The 27B ternary model file is 8.2 GB and fits on one 12 GB NVIDIA GPU","excerpt":"The whole model, 8.2 GB (1.6 bits per ternary weight), fits on one 12 GB NVIDIA GPU.","status":"not_checked"},{"kind":"specific","claim":"Reports 34.8 tokens/s decode for the 27B adapter model on an RTX 3090 Ti","excerpt":"34.8 tok/s","status":"not_checked"},{"kind":"specific","claim":"Reports GSM8K score of 0.895 for the 27B model, equal to full precision","excerpt":"GSM8K 0.895 (full precision 0.895)","status":"not_checked"},{"kind":"specific","claim":"Its export is exact: next-token loss in llama.cpp matches the trainer's forward pass","excerpt":"The export is exact: next-token loss in llama.cpp matches the trainer's own forward.","status":"not_checked"}],"alternatives":[],"written_by":"AI, from the project's README","date":"2026-11-12"},"why_now":null,"owner_supplied":null,"score":{"verdict":"overhyped","hype":90,"reality":39,"gap":51,"momentum":null,"confidence":60,"version":"score-2.0","rule":"verdict-2.0","calculated_at":"2026-11-22T07:00:00+00:00","verdict_source":"rule"},"early_read":null,"changes":[{"date":"2026-11-09T13:25:07+00:00","text":"Latest release v0.1.0","source":"https://github.com/oguzeldereli/ttf-home/releases"},{"date":"2026-11-09T13:24:37+00:00","text":"Repository created","source":"https://github.com/oguzeldereli/ttf-home"}],"cite":"ttf-home is a GitHub project at https://github.com/oguzeldereli/ttf-home written mainly in C++, described there as \"Ternary language models on a home computer: run, train, convert, fine-tune, export to GGUF\".","attribution":"Data from Mid or Real, https://midorreal.com/project/oguzeldereli-ttf-home, checked 22 November 2026. Attribution with a link is required.","still_real":null}