Ternary-Bonsai-1.7B-gguf

Model Card for Ternary Bonsai 1.7B (GGUF, re-host)

Prism ML's Ternary Bonsai 1.7B Q2_0_g128 weights, re-hosted 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. The weights are upstream's, unchanged; the GGUF header carries one added key (below). 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

The tensor data is byte-for-byte identical to upstream's file. Only the metadata header differs: one key, general.generation_config_json, holds a sampling card as text so the per-mode preset matrix travels inside the weights instead of in a sidecar file that can be lost.

upstream Ternary-Bonsai-1.7B-Q2_0.ggxf
  sha256 d97d94eb564590c9f0300e54d3f87bbbb25a78693d0ade9f6e177973dcb8228a
  463,290,464 bytes, 310 tensors, 35 KV
this file
  463,292,000 bytes, 310 tensors, 36 KV
  tensor data region identical (verified with cmp across the two
  data offsets, 457,345,184 bytes)

Nothing was re-quantized or converted. To reproduce the difference, take the upstream file and rewrite its header; the tensor region copies verbatim.

The card's per-mode values come from upstream's own general.sampling.* keys (temperature 0.5, top_p 0.85, top_k 20, min_p 0.0), which the file already carried and which this key leaves untouched. The card adds the rows those scalars cannot express: a presence penalty of 1.5 in the non-thinking modes, this lineage's policy for curbing small-model repetition.

Model Details

Model Description

Ternary Bonsai 1.7B is a dense Qwen3 model (general.architecture = qwen3): grouped-query attention (GQA) with a SwiGLU MLP, RoPE, and RMSNorm across every layer -- no Gated DeltaNet / linear-attention blocks (those belong to the 27B hybrid). The weights are quantized to a true ternary alphabet {-1, 0, +1} (~1.58 bits of information per weight). This is the Q2_0_g128 build: each 128-weight block is { FP16 scale d; 2-bit codes qs[32] } (34 bytes, 2.125 bits/weight effective) and dequantizes as w = (code - 1) * d.

• Developed by: Prism ML (model + ternary quantization), built from Qwen3-1.7B by Alibaba Cloud; re-hosted by leok7v
• Model type: Dense grouped-query-attention causal language model, ternary-quantized GGUF (llama.cpp)
• 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-1.7B-gguf

Model Sources

• Repository (this re-host): https://huggingface.co/leok7v/Ternary-Bonsai-1.7B-gguf
• Original (authoritative) repository: prism-ml/Ternary-Bonsai-1.7B-gguf (Apache 2.0)
• Base model: Qwen/Qwen3-1.7B (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.

Copyright 2026-present Prism ML, Inc.; built from Qwen3-1.7B, Copyright 2024 Alibaba Cloud. Apache 2.0, inherited unchanged from upstream; LICENSE and NOTICE.txt are included here as upstream ships them.

Contents

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.