Prism ML Website | White Paper | Demo & Examples | Discord
Ternary-Bonsai-4B-mlx-2bit
Ternary (1.58-bit) language model for Apple Silicon
7.1x smaller than FP16 | 4.8x faster on M4 Pro | 50 tok/s on iPhone | runs on Mac, iPhone, iPad
Highlights
- 1.05 GiB (1.13 GB) packed 2-bit size (down from 8.04 GB FP16) — runs on any Mac or iPhone
- Ternary weights {-1, 0, +1} across embeddings, attention projections, MLP projections, and LM head
- 70.7 avg benchmark score across 6 categories — competitive with full-precision 4B models
- MLX-native format with group size 128 and FP16 scaling
Resources
- White Paper
- Demo repo — examples for serving, benchmarking, and integrating Bonsai
- Discord — community support and updates
- Kernels: MLX (Apple Silicon) · mlx-swift (iOS/macOS) — 2-bit format is supported out of the box
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 | Ternary g128: {-1, 0, +1} with FP16 group-wise scaling |
| Packed 2-bit size | 1.05 GiB (1.13 GB) |
| Ternary coverage | Embeddings, attention projections, MLP projections, LM head |
| License | Apache 2.0 |
Quantization Format: Ternary 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}
The information-theoretic cost is log2(3) ≈ 1.585 bits per weight, plus FP16 group scales (16 bits per 128 weights), for a theoretical minimum of ~1.71 bits/weight. This release uses the MLX 2-bit format, which stores each ternary value in 2 bits plus group scales, for an effective ~2.125 bits/weight.
Memory
| Format | Size | Reduction | Ratio |
|---|---|---|---|
| FP16 | 8.04 GB | -- | 1.0x |
| MLX 2-bit g128 | 1.05 GiB (1.13 GB) | 85.9% | 7.1x |
Quickstart
MLX (Python)
pip install mlx-lm
from mlx_lm import load, generate
model, tokenizer = load("prism-ml/Ternary-Bonsai-4B-mlx-2bit")
response = generate(
model,
tokenizer,
prompt="Explain quantum computing in simple terms.",
max_tokens=256,
)
print(response)
Throughput (MLX / Apple Silicon)
| Platform | Backend | PP512 (tok/s) | TG128 (tok/s) | FP16 TG (tok/s) | Speedup |
|---|---|---|---|---|---|
| M4 Pro 48 GB | MLX (Python) | 817 | 133 | 28 | 4.8x |
iPhone 17 Pro Max (MLX Swift)
| Platform | Backend | PP512 (tok/s) | TG128 (tok/s) | 4-bit TG (tok/s) | Speedup |
|---|---|---|---|---|---|
| iPhone 17 Pro Max | MLX Swift | 659 | 50 | 27 | 1.8x |
Benchmarks
Evaluated with EvalScope v1.4.2 + vLLM 0.15.1 on NVIDIA H100. Full benchmark suite (10 benchmarks):
| Model | Size | Avg | MMLU-R | MuSR | IFEval | GSM8K | HE+ | BFCLv3 |
|---|---|---|---|---|---|---|---|---|
| Ternary Bonsai 4B | 0.86 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 | 0.86 GB | 1.426 |
| 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