Prism ML Website | White Paper | Demo & Examples | Discord

Ternary-Bonsai-4B-gguf

Ternary (1.58-bit) language model in GGUF Q2_0 format for llama.cpp

Resources

White Paper
Demo repo — examples for serving, benchmarking, and integrating Bonsai
Discord — community support and updates
Kernels: Q2_0 is not yet in mainline llama.cpp. Use our fork at PrismML-Eng/llama.cpp (prism branch, default) which adds Q2_0 support for CPU (NEON/generic) and Metal. Upstream PR coming soon.

Model Overview

Item	Specification
Base model	Qwen3-4B
Parameters	4.0B (~3.6B non-embedding)
Architecture	GQA (32 query / 8 KV heads), SwiGLU MLP, RoPE, RMSNorm
Layers	36 Transformer decoder blocks
Context length	32,768 tokens
Vocab size	151,936
Weight format	GGUF Q2_0 g128: {-1, 0, +1} with FP16 group-wise scaling
Packed Q2_0 size	1,020 MiB (1.07 GB)
Ternary coverage	Embeddings, attention projections, MLP projections, LM head
License	Apache 2.0

Quantization Format: GGUF Q2_0 (g128)

Each weight takes a value from {-1, 0, +1}, with one shared FP16 scale per group of 128 weights:

w_i = scale_g * t_i,    t_i in {-1, 0, +1}

Q2_0 encodes each weight as a 2-bit code q in {0, 1, 2, 3}, dequantized via w = (q - 1) * scale. One 128-element block is 34 bytes (2 bytes FP16 scale + 32 bytes of packed 2-bit codes) for an effective 2.125 bits/weight. The fourth code point (q = 3, reconstructing to +2 * scale) is reserved for future extensions; for ternary weights it is unused.

Memory

Format	Size	Reduction	Ratio
FP16	8.04 GB	--	1.0x
GGUF Q2_0 g128	1,020 MiB (1.07 GB)	86.3%	7.3x

Files in this repo

File	Format	Size	Recommended
`Ternary-Bonsai-4B-F16.gguf`	FP16	8.04 GB	baseline / re-quantization source
`Ternary-Bonsai-4B-Q2_0.gguf`	Q2_0 (g128)	1,020 MB	recommended (lossless for ternary)

Quickstart

Build from the Prism fork

git clone https://github.com/PrismML-Eng/llama.cpp
cd llama.cpp
cmake -B build -DGGML_METAL=ON   # or -DGGML_CUDA=ON, -DGGML_VULKAN=ON
cmake --build build -j

`llama.cpp` CLI

./build/bin/llama-cli \
  -m Ternary-Bonsai-4B-Q2_0.gguf \
  -p "Explain quantum computing in simple terms." \
  -n 256

`llama.cpp` server

./build/bin/llama-server -m Ternary-Bonsai-4B-Q2_0.gguf -c 4096

Throughput (llama.cpp, Apple M4 Pro 48 GB)

Backend	PP512 (tok/s)	TG128 (tok/s)
Metal (GPU)	826	120
NEON CPU (10 t)	226	56

Flags: -ngl 99 -fa 1 for Metal; -ngl 0 -fa 1 -t 10 for CPU.

Fidelity (Q2_0 vs FP16 baseline)

Q2_0 is effectively lossless for ternary weights — the ternary values land exactly on three of the four 2-bit code points, so quantize/dequantize is bit-exact in the absence of FP16 scale rounding.

Benchmarks

Evaluated with EvalScope v1.4.2 + vLLM 0.15.1 on NVIDIA H100. Full benchmark suite:

Model	Size	Avg	MMLU-R	MuSR	IFEval	GSM8K	HE+	BFCLv3
Ternary Bonsai 4B	1.02 GB	70.7	69.7	45.1	72.1	90.5	78.7	67.8
1-bit Bonsai 4B (prior)	0.57 GB	62.7	58.7	41.4	69.6	87.3	71.3	48.0
Qwen 3 4B	8.04 GB	77.1	79.8	57.4	80.0	92.1	74.4	78.9
Ministral3 3B	6.86 GB	73.2	77.5	56.5	73.1	91.4	69.5	71.3
Gemma 3 4B	7.76 GB	67.9	66.0	46.3	73.0	89.8	67.1	65.1
Llama 3.2 3B	6.43 GB	64.4	65.5	48.9	78.3	80.1	52.4	60.9

Intelligence Density

density = -ln(1 - score/100) / size_GB

Model	Size	Intelligence Density (1/GB)
Ternary Bonsai 4B	1.02 GB	1.202
1-bit Bonsai 4B (prior)	0.57 GB	1.744
Ministral3 3B	6.86 GB	0.192
Qwen 3 4B	8.04 GB	0.183
Llama 3.2 3B	6.43 GB	0.161
Gemma 3 4B	7.76 GB	0.146

Citation

@techreport{ternarybonsai,
    title   = {Ternary Bonsai: 1.58-bit Language Models at 8B, 4B, and 1.7B Scale},
    author  = {Prism ML},
    year    = {2026},
    month   = {April},
    url     = {https://prismml.com}
}

Contact

For questions, feedback, or collaboration inquiries: contact@prismml.com

prism-ml/Ternary-Bonsai-4B-gguf