Ternary Bonsai 2 27B

26.9B parameter open-weight model. A ternary re-training of Qwen3.8-27B by PrismML: every language weight is -1, 0 or +1 with FP16 scales per group of 128 (embeddings, attention, MLP and the output head), on the base model's hybrid-attention architecture (about 75% linear attention), with a 262,144-token context and an optional vision projector (model card read 2026-10-23). 26.9B parameters in the GGUF metadata; the card's 27.36B adds the 0.46B vision tower. Apache 2.0. The vendor publishes only packed GGUF and MLX releases, so the GGUF repository is the release itself, not a re-upload. Two packings: PTQ1_0 (1.75 bits per weight, 5.95 GB) and PQ2_0 (2.13 bits per weight, 7.21 GB). Neither is a Q4_K_M: the figure here is the PTQ1_0 file. Both need PrismML's llama.cpp fork (github.com/PrismML-Eng/llama.cpp); stock Ollama, LM Studio and upstream llama.cpp do not read them, so no Ollama library tag exists (registry checked 2026-10-23) and the Checker does not list it. Measured on this project's own hardware with the fork at its pinned release tag prism-b10777-e3de715, an RTX 3080 (10 GB) with nothing else on the card: entirely in VRAM, PTQ1_0 generated 78.4 tok/s and used 6.75 GB at a 4,096 context and 8.57 GB at 32,768; PQ2_0 generated 66.8 tok/s and used 7.90 GB and 9.72 GB; prompt processing ran at about 1,290 tok/s for both, and both answered a short reasoning prompt and a coding task correctly (2026-10-23, docs/lab/radar-2026-10-23-model-validation.md). Plan around 7 GB of free VRAM for PTQ1_0 at a short context.