Ternary Bonsai 2 27B

A 27B multimodal model that fits in about 6GB, the size of a couple of movies. The trick is in how numbers get stored: instead of full-precision weights, every weight is just -1, 0, or +1, which packs a whole model into a download you can actually finish on hotel wifi. It sees images, reads 262k tokens of context (a novel series, not a chapter), and keeps 98.2 percent of its full precision quality on a 20 benchmark suite. Run it with llama.cpp or MLX. If you ever wanted a desktop model that does vision, long documents, and reasoning without a GPU budget, this is the current default answer.