{"name":"ttf-ternary","repo":"oguzeldereli/ttf-ternary","url":"https://midorreal.com/project/oguzeldereli-ttf-ternary","repository":"https://github.com/oguzeldereli/ttf-ternary","description":"Two-timescale threshold flips (TTF): training and converting ternary {-1,0,1} LLMs without full-precision master weights. ~1 byte of optimizer state per weight; 23B trains on a 16 GB GPU, Qwen3.8-27B converted and recovered on one 24 GB card.","facts":{"language":{"value":"Python","source":"https://github.com/oguzeldereli/ttf-ternary","checked":"22 Nov 2026"},"last_release":{"value":"v0.3.0, 9 Nov 2026","source":"https://github.com/oguzeldereli/ttf-ternary/releases","checked":"22 Nov 2026"},"releases":{"value":"3","source":"https://github.com/oguzeldereli/ttf-ternary/releases","checked":"21 Nov 2026"},"contributors":{"value":"1","source":"https://github.com/oguzeldereli/ttf-ternary/graphs/contributors","checked":"21 Nov 2026"},"archived":{"value":"no","source":"https://github.com/oguzeldereli/ttf-ternary","checked":"22 Nov 2026"}},"stars":50.0,"writeup":{"what_it_is":"ttf-ternary is a training framework for ternary language models with values in {-1, 0, +1}. It trains and converts these models without full-precision master weights, using a two-timescale threshold flip (TTF) method and low-rank optimizer state.","audience":"Researchers and developers training or converting ternary language models on NVIDIA GPUs","claims":[{"kind":"specific","claim":"Trains ternary models without full-precision master weights, keeping weights packed at 1.6 bits","excerpt":"The weights stay packed at 1.6\nbits (5 trits per byte) for the whole of training.","status":"not_checked"},{"kind":"specific","claim":"Optimizer state is about one byte per weight","excerpt":"The optimizer state is low-rank, about one byte per weight in place\nof the 12-16 bytes of master weights plus Adam.","status":"not_checked"},{"kind":"specific","claim":"A 23B-parameter model trains on one 16 GB GPU","excerpt":"A 23B-parameter model trains on one 16 GB GPU, and a converted\nQwen3.8-27B trains on one 24 GB card.","status":"not_checked"},{"kind":"specific","claim":"Reports validation results on Wikipedia and FineWeb-Edu compared with BitNet b1.58 and full precision","excerpt":"At 600M tokens of FineWeb-Edu TTF leads master + extras by 0.056 (3.0423 vs 3.0982).","status":"not_checked"},{"kind":"specific","claim":"Has fast Triton kernels that use int8 tensor cores","excerpt":"The kernels are Triton and use the int8 tensor cores","status":"not_checked"}],"alternatives":[],"written_by":"AI, from the project's README","date":"2026-11-12"},"why_now":null,"owner_supplied":null,"score":{"verdict":"mid","hype":85,"reality":60,"gap":25,"momentum":null,"confidence":60,"version":"score-2.0","rule":"verdict-2.0","calculated_at":"2026-11-22T08:00:00+00:00","verdict_source":"rule"},"early_read":null,"changes":[{"date":"2026-11-09T13:24:07+00:00","text":"Latest release v0.3.0","source":"https://github.com/oguzeldereli/ttf-ternary/releases"},{"date":"2026-10-15T12:03:08+00:00","text":"First release","source":"https://github.com/oguzeldereli/ttf-ternary/releases"},{"date":"2026-08-31T16:04:28+00:00","text":"Repository created","source":"https://github.com/oguzeldereli/ttf-ternary"}],"cite":"ttf-ternary is a GitHub project at https://github.com/oguzeldereli/ttf-ternary written mainly in Python, described there as \"Two-timescale threshold flips (TTF): training and converting ternary {-1,0,1} LLMs without full-precision master weights. ~1 byte of optimizer state per weight; 23B trains on a 16 GB GPU, Qwen3.8-27B converted and recovered on one 24 GB card\".","attribution":"Data from Mid or Real, https://midorreal.com/project/oguzeldereli-ttf-ternary, checked 22 November 2026. Attribution with a link is required.","still_real":null}