Prism ML's Ternary Bonsai 27B Q2_0_g128 weights and their Q8_0 vision projector, re-hosted as ONE GGUF under the Apache 2.0 license, as a stable pinned snapshot that survives if the upstream repo moves or is removed. This is NOT the authoritative source. Every tensor is upstream's, unchanged. For the full model card, benchmarks, other quants, and the whitepaper, use the original repository.

File format

.ggxf is GGUF plus the tensors and metadata keys this project adds, so these files load in Gadeon and are not expected to load in other GGUF readers.

What differs from upstream

Upstream ships the text model and the vision projector as two files, plus no sampling card. This repo ships one file that carries both tensor sets and the card, so a model is one file: nothing beside the weights can go missing and leave the model quietly text-only or quietly mis-sampled.

upstream Ternary-Bonsai-27B-Q2_0.ggxf
  sha256 868c11714cf8fe47f5ec9eeb2be0ab1a337112886f92ee0ede6b855c4fa31757
  7,165,121,600 bytes, 851 tensors, 37 KV
upstream Ternary-Bonsai-27B-mmproj-Q8_0.ggxf
  sha256 eb561d41a7bbeb0fcf04883c8af11078ef6cae0a66862a0b68443cfca495269d
  629,246,880 bytes, 334 tensors, 24 KV
this file
  7,794,369,344 bytes, 1185 tensors, 53 KV
  851 text tensors then 334 `v.*` / `mm.*` projector tensors,
  the projector's `clip.*` keys, and `general.generation_config_json`

Both tensor data regions were compared byte for byte after the merge and are identical to their upstream sources (text 7,154,128,896 bytes, projector 629,227,328 bytes). Nothing was re-quantized or converted.

The card's per-mode values come from upstream's own general.sampling.* keys (temperature 1.0, top_p 0.95, top_k 20), which the file already carried and which this key leaves untouched. The card adds the rows those scalars cannot express: a lower temperature for vision turns, and the non-thinking row's temperature 0.7 / top_p 0.8 / presence penalty 1.5.

Compatibility

Mainline llama.cpp will NOT load this file. Its loader requires the model file to hold exactly the model's own tensors and takes the projector separately via --mmproj, so the 334 extra tensors here make it fail with wrong number of tensors. Use the upstream two-file repo for llama.cpp. This layout is for runtimes that read the projector out of the same file.

Model Details

Model Description

Ternary Bonsai 27B is a Qwen3.5 hybrid model (general.architecture = qwen35): of its 64 layers, 48 are Gated DeltaNet linear-attention (state-space) blocks and 16 are full softmax-attention blocks, interleaved "three linear, one attention". The weights are quantized to a true ternary alphabet {-1, 0, +1} at ~1.71 bits per weight. This is the Q2_0_g128 build: each 128-weight block is { FP16 scale d; 2-bit codes qs[32] } and dequantizes as w = (code - 1) * d. The vision projector is a 27-block Qwen3-VL tower at Q8_0.

Developed by: Prism ML (model + ternary quantization), built from Qwen3.6-27B by Alibaba Cloud; re-hosted by leok7v
Model type: Hybrid linear/softmax-attention causal language model with a vision projector, ternary-quantized GGUF
Language(s): English and the languages of the base model
License: Apache 2.0 (inherited unchanged from the upstream model)
Derived from model: prism-ml/Ternary-Bonsai-27B-gguf

Model Sources

Repository (this re-host): https://huggingface.co/leok7v/Ternary-Bonsai-27B-gguf
Original (authoritative) repository: prism-ml/Ternary-Bonsai-27B-gguf (Apache 2.0)
Q2_0 kernels: PrismML-Eng/llama.cpp (prism branch) -- Q2_0 is not yet in mainline llama.cpp.

Attribution

Created using Bonsai by Prism ML.

File	Notes
`Ternary-Bonsai-27B-Q2_0.ggxf` (~7.8 GB)	Upstream text weights and vision projector in one file, header carrying the sampling card.
`LICENSE.txt`, `NOTICE.txt`	Upstream's, unchanged.

Why a re-host and not a fork

Hugging Face does not support forking a model repository, so re-hosting the weight file is the only way to keep a pinned snapshot; the storage and bandwidth are this repo's cost, not upstream's. If you just want the model, prefer the original.

leok7v/Ternary-Bonsai-27B-gguf

Model Card for Ternary Bonsai 27B (GGUF, re-host, single file)