---

license: apache-2.0

base_model: prism-ml/Ternary-Bonsai-1.7B-gguf

tags:

• gguf
• quantized
• q2_0
• ternary

---

AJAN-SIMIT Ternary-Bonsai Q2_0 (GGUF)

Q2_0 GGUF quantizations of Prism ML's Ternary-Bonsai model family, produced by DarkStar

Initiative for the Ajan Simit offline

mobile AI project (an on-device Android voice assistant).

We did not train or create the base model. This repository contains only our own

re-quantization of Prism ML's publicly released weights, done so the model is small enough to

run entirely on-device on a phone.

Provenance / credits

• Base model: Qwen3-1.7B (Apache 2.0, Qwen team)
• Ternary fine-tune: prism-ml/Ternary-Bonsai (Apache 2.0, Prism ML)
• This Q2_0 re-quantization: DarkStar Initiative / Ajan Simit, produced with a self-built

llama-quantize from Prism ML's own fork, PrismML-Eng/llama.cpp

(prism branch) -- credited per their own model card's request.

• License: Apache 2.0, inherited unchanged from the base model and the fine-tune. No added

restrictions. A full copy of the license text is included in this repo as LICENSE.

If you use the original Ternary-Bonsai model itself (not just this re-quantization), please

cite Prism ML directly:

@techreport{ternarybonsai,
  title  = {Ternary Bonsai: 1.58-bit Language Models},
  author = {Prism ML},
}

What's different from the upstream Prism ML release

Only the quantization level and the file/repo naming:

• Quantization: Q2_0 (2.125 bpw), quantized from Prism ML's F16 release. No retraining, no

architecture changes, no fine-tuning of our own.

• Naming: the AJAN-SIMIT- prefix marks these files as our own community re-quantization

and build, produced for the Ajan Simit app specifically -- not an official Prism ML release,

and not endorsed by Prism ML or the Qwen team.

Files

| File | Size |

|---|---:|

| AJAN-SIMIT-Ternary-Bonsai-1.7B-Q2_0.gguf | ~463 MB |

| AJAN-SIMIT-Ternary-Bonsai-4B-Q2_0.gguf | ~1.07 GB |

| AJAN-SIMIT-Ternary-Bonsai-8B-Q2_0.gguf | ~2.18 GB |

| AJAN-SIMIT-Ternary-Bonsai-27B-Q2_0.gguf | ~8.25 GB |

Usage

Compatible with any recent llama.cpp-based inference engine. Q2_0 is a first-class quant

type in Prism ML's llama.cpp fork; mainline llama.cpp support may vary by version.