---
license: apache-2.0
library_name: llama.cpp
pipeline_tag: text-generation
base_model:
  - prism-ml/Ternary-Bonsai-27B-gguf
  - lym00/Qwen3.6-27B-MTP-ONLY-GGUF
tags:
  - gguf
  - llama.cpp
  - ternary
  - tq2_0
  - mtp
  - nextn
  - speculative-decoding
  - self-contained
---

# Ternary Bonsai 27B — One-File MTP GGUF for stock llama.cpp

A self-contained Ternary Bonsai 27B GGUF with one embedded Qwen-style NextN/MTP layer.

The same file supports ordinary inference and `draft-mtp`. No separate draft model, sidecar GGUF, or custom llama.cpp fork is required.

This is an independent community release. It is not an official Prism ML, Qwen, Alibaba Cloud, or llama.cpp release, and no upstream party endorses it.

## Model file

| File | Bytes | SHA-256 |
|---|---:|---|
| `Ternary-Bonsai-27B-MTP-TQ2_0.gguf` | 8,785,215,776 | `480c695b0de5b78b720d7f500eb3e1650359d0ac8e18c29b231a14b2d9c9bf20` |

## Use with llama.cpp

A stock llama.cpp build with `qwen35` and `draft-mtp` support is required.

Ordinary inference:

```bash
llama-cli \
  -m Ternary-Bonsai-27B-MTP-TQ2_0.gguf \
  -p "Write a concise explanation of ternary neural-network weights." \
  -n 128
```

MTP speculative decoding from the same GGUF:

```bash
llama-cli \
  -m Ternary-Bonsai-27B-MTP-TQ2_0.gguf \
  --spec-type draft-mtp \
  -p "Write a concise explanation of ternary neural-network weights." \
  -n 128
```

No `-md` argument is needed.

## Download and verify

```bash
hf download vinpix/Ternary-Bonsai-27B-Stock-MTP-GGUF \
  Ternary-Bonsai-27B-MTP-TQ2_0.gguf \
  SHA256SUMS \
  --local-dir .
```

Verify the downloaded payload:

```bash
sha256sum -c SHA256SUMS --ignore-missing
```

Expected model digest:

```text
480c695b0de5b78b720d7f500eb3e1650359d0ac8e18c29b231a14b2d9c9bf20
```

## Portable artifact facts

These properties are independent of processor, accelerator, operating system, thread count, memory bandwidth, and runtime timing:

| Property | Value |
|---|---|
| GGUF version | 3 |
| Architecture | `qwen35` |
| File size | 8,785,215,776 bytes |
| Tensor count | 866 |
| Metadata key/value count | 42 |
| Transformer blocks | 65 |
| Embedded NextN/MTP layers | 1 |
| Main low-bit tensor type | `TQ2_0` |
| Tensor-type inventory | F32 359; Q2_K 17; Q4_K 1; Q5_K 6; Q6_K 3; TQ2_0 480 |

The public GGUF was checked for exact length, SHA-256, parseable GGUF v3 metadata, tensor inventory, block inventory, and absence of local absolute paths in structured metadata.

Build-system source paths in `quantize.imatrix.file` and `quantize.imatrix.dataset` were replaced with neutral public values. Tensor data and tensor offsets were not modified.

No device-specific throughput, prefill rate, memory use, hostname, mount point, username, or local filesystem path is published in this repository.

## Construction and provenance

The release combines:

- Ternary Bonsai 27B trunk from `prism-ml/Ternary-Bonsai-27B-gguf`
- Qwen3.6-derived MTP-only material packaged by `lym00/Qwen3.6-27B-MTP-ONLY-GGUF`

Pinned construction sources:

| Role | Source revision | Source SHA-256 |
|---|---|---|
| Ternary Bonsai 27B F16 trunk | `3f8cc399dde45ac0475d023634974407af34907c` | `f659ca3dd7e28ada5d8b5f3637862d0d51ef433bde032ec4c8990ed27c91a385` |
| Qwen3.6 27B MTP-only BF16 package | `03b35db648da71e23ecfb239f70661e069b16054` | `97697fc5278d4bfd0afaa733ba28c4338cb0396b98d5f5acb781e515f72b52c0` |

The donor package identifies the underlying model as `Qwen/Qwen3.6-27B`, but it does not pin the underlying Qwen source revision. This release does not infer or substitute an unrecorded revision.

## Scope

Self-contained MTP GGUFs already exist for Qwen-family models. The distinguishing feature of this release is the combination of a Ternary Bonsai 27B trunk and a functional embedded NextN/MTP predictor in one stock-compatible GGUF.

No universal priority claim is made.

## Limitations

- The NextN/MTP layer was grafted from a compatible Qwen3.6-derived donor; it was not jointly trained with the Bonsai trunk.
- MTP acceptance and speedup depend on prompt, context, sampling, backend, runtime revision, and hardware.
- Aggressive quantization can reduce quality relative to higher-precision releases.
- Generated output should be reviewed before use in consequential settings.
- No claim of statistical superiority over other Bonsai or Qwen releases is made.

## License and attribution

Released under the Apache License 2.0. See `LICENSE` and `NOTICE`.

Upstream attribution is preserved for Prism ML, Qwen/Alibaba Cloud, and the pinned MTP-only package by lym00.