# Upstreaming Prism ternary types 142/143 into llama.cpp and Ollama > Prism's Bonsai-demo README says Bonsai 2 support is being upstreamed in smaller PRs that target the official `Q2_0` format and that Bonsai 2 still requires the PrismML fork. Parent: [Mac local LLMs: Quantization formats and methods](https://llms-explorer.com/tree/mac-local-llms-quantization-formats-and-methods/) · 1 facets · 16 facts · page: https://llms-explorer.com/tree/upstreaming-prism-ternary-types-142-143-into-lla/ ## Facts - Prism's Bonsai-demo README says Bonsai 2 support is being upstreamed in smaller PRs that target the official `Q2_0` format and that Bonsai 2 still requires the PrismML fork. — [source](https://github.com/PrismML-Eng/Bonsai-demo) - The README status table, checked 2026-09-25, lists llama.cpp PRs 27779 and 29094 and 29095 as merged, 29096 and 29243 as open, and 29100 and 29101 as draft. — [source](https://github.com/PrismML-Eng/Bonsai-demo) - The README says PQ2_0 and PTQ1_0 "remain fork-specific packings" and that the upstream effort "does not promise support for those types". — [source](https://github.com/PrismML-Eng/Bonsai-demo) - llama.cpp PR 29094 makes the Metal FWHT kernel accept an F16 source and gates `supports_op` to the four kernel sizes. — [source](https://github.com/ggml-org/llama.cpp/pull/29094) - llama.cpp PR 29096 does the same for CUDA, with `test-backend-ops -o MUL_MAT_HADAMARD` 15/15. — [source](https://github.com/ggml-org/llama.cpp/pull/29096) - Prism's formats page says Ternary Bonsai 2 needs an activation-side Walsh-Hadamard transform that is "not upstream yet", so every backend needs the fork binaries `prism-b10658` or newer. — [source](https://docs.prismml.com/download/formats) - The formats page says Ollama and other tools that bundle stock llama.cpp cannot run Ternary Bonsai 2. — [source](https://docs.prismml.com/download/formats) - The formats page says the MLX package for Ternary Bonsai 2 uses a 2-bit group-128 packing analogous to PQ2_0 and runs on stock MLX. — [source](https://docs.prismml.com/download/formats) - A llama.cpp member commented on issue 29058 on 2026-09-19 with only "See also #22019". — [source](https://github.com/ggml-org/llama.cpp/issues/29058) - An issue 29058 commenter reports master's `qwen35.cpp` MTP path reads `token_embd` with a raw `ggml_get_rows`, and a folded model with `--spec-type draft-mtp` fails with `Hadamard-latent table 'token_embd.weight' is read without the inverse transform`. — [source](https://github.com/ggml-org/llama.cpp/issues/29058) - The same commenter says the Prism fork fixed this in commit 422590f5 on 2026-09-21. — [source](https://github.com/ggml-org/llama.cpp/issues/29058) - The same commenter measured a standalone MTP sidecar at 5.16 t/s against 16.6 t/s without speculation, with rho 0.43 versus about 0.06 for a shared table. — [source](https://github.com/ggml-org/llama.cpp/issues/29058) - The same commenter measured MTP acceptance on the 27B pack at 65.8%, 52.2%, 67.9%, 82.3% and 84.1% for 8k, 32k, 64k, 128k and 191k tokens of context. — [source](https://github.com/ggml-org/llama.cpp/issues/29058) - The same commenter says folding the terminal norm gain `gf / rms(gf)` into `eh_proj` raised 0.8B acceptance from 10.7% to 28.4% and 27B from 35.6% to 40.5%. — [source](https://github.com/ggml-org/llama.cpp/issues/29058) - The same commenter says an m=3 file can be served as m=2 without requantizing by dropping a plane. — [source](https://github.com/ggml-org/llama.cpp/issues/29058) - The Prism formats page lists PTQ1_0 at 1.76 bpw and 5.93 GB and PQ2_0 at 2.16 bpw and 7.25 GB, which differs from the Hugging Face card's 1.75 bpw, 5.95 GB and 2.13 bpw, 7.21 GB. — [source](https://docs.prismml.com/download/formats)