Ternary Bonsai 2 9B

9B parameter open-weight model. A ternary version of Qwen3.5-9B by PrismML: every language weight is -1, 0 or +1 with FP16 scales per group of 128, on the base model's unchanged hybrid-attention architecture (32 blocks, hidden size 4,096, 24 linear-attention and 8 full-attention layers), with a 262,144-token context and an optional 0.46B vision tower shipped as a separate projector (model card read 2026-11-03). 8.95B parameters in the language model (GGUF header); the Hugging Face API's GGUF summary reports 1.47B because it reads the repository's DFlash draft file instead. Apache 2.0. The vendor publishes only packed GGUF and MLX releases, so the GGUF repository is the release itself, not a re-upload. Three packings of the same weights: PTQ1_0 (2.00 GB file), PQ2_0 (2.42 GB) and Q2_0 (2.56 GB). Only Q2_0 runs outside PrismML's llama.cpp fork (github.com/PrismML-Eng/llama.cpp): upstream llama.cpp b11847 and Ollama 0.40.10 load it, and Ollama's library carries it as bonsai-2:9b-q2_0, whose model layer is byte-identical to the Hugging Face Q2_0 file (sha256 1f884aa56aac...) and adds a 0.62 GB vision projector, 3.2 GB in all (registry manifest read 2026-11-03; `ollama show` lists a minimum of Ollama 0.40.6). PTQ1_0 and PQ2_0 need the fork, and Ollama rejects a PTQ1_0 import. The slot here is the Q2_0 file, not a Q4_K_M. Measured on this project's own hardware (docs/lab/radar-2026-10-30-model-validation.md), an RTX 3080 (10 GB) with nothing else on the card: the Q2_0 file generated 167 tok/s in upstream llama.cpp and 153 tok/s in Ollama, and PTQ1_0 and PQ2_0 through the fork generated 196 and 172 tok/s, all entirely in VRAM; the server held 2,538 to 3,018 MiB at a 4,096 context and 3,462 to 3,942 MiB at 32,768. Beside a running camera-recording service on the same card, the library tag generated about 80 tok/s and took about 7,200 MiB of the card in total (2026-11-03). On an RTX 4070 Ti desktop the three packings generated 127 to 138 tok/s through the fork. A three-sentence answer and a coding task were correct on all three packings. The model reasons by default and can spend thousands of tokens on a short prompt: the vendor advises at least 8,192 output tokens with thinking on, or turning thinking off.