Four dominant threads in the discussion
| # | Theme | Supporting quotations |
|---|---|---|
| 1 | Quantization efficiency – bits vs. trits | “Ternary Bonsai 27B uses ternary {−1, 0, +1} weights … giving a true 1.71 effective bits per weight.” – NitpickLawyer “5 trits (243 states) into a byte gives 1.6 bits per trit.” – petu |
| 2 | Running the models on consumer hardware | “One of the links … requires an iPhone 17 Pro or Pro Max to run the 27B model.” – smallerize “Doing some naive math, the F16 filesize is ~53.8 GB … the 1‑bit version is ~3.8 GB, about 7 % of the original size.” – Catloafdev |
| 3 | Performance trade‑offs & reasoning‑loop issues | “‘The model gets stuck in a reasoning loop … much less often than 27B in my experience.’” – dofm “It fails the ‘Jabberwocky’ test.” – raylad |
| 4 | Business & strategic implications | “If you read to the bottom of the page, it says they’re funded by a few people, and one of them is Samsung.” – trollbridge “Apple would punish him severely unless they cleared it in advance …” – CharlesW |
These four themes capture the bulk of the conversation: the mathematics of ultra‑compact representations, the practical hurdles of deploying such models on phones or modest GPUs, the observed strengths and failure modes of the quantized LLMs, and the broader industry ramifications for hardware makers and open‑source strategy.