GoKawiilToday, we’re announcing Ternary Bonsai, a new family of 1.58-bit language models designed to balance strict memory constraints with high accuracy requirements.
This release builds on the efficiency frontier we began exploring with the recently released 1-bit Bonsai models. The 1-bit family showed that extreme compression could still produce commercially useful language models. Ternary Bonsai targets a different point on that curve: a modest increase in size for a meaningful gain in performance.
The models are available in three sizes: 8B, 4B, and 1.7B parameters. By using ternary weights {-1, 0, +1}, these models achieve a memory footprint approximately 9x smaller than standard 16-bit models while outperforming most peers in their respective parameter classes on standard benchmarks.
A true ternary model
Ternary Bonsai implements 1.58-bit representation throughout the entire network architecture. There are no higher-precision escape hatches. Embeddings, attention layers, MLPs, and the LM head all use the same 1.58-bit representation.
The models employ a group-wise quantization scheme in which each weight is constrained to one of three values: {-s, 0, +s}. These three states are encoded as (-1, 0, +1) using 1.58 bits per weight, together with a shared FP16 scale factor (s) for each group of 128 weights.
Benchmark performance
Compared to the 1-bit Bonsai 8B, the Ternary Bonsai 8B scores 5 points higher on average across benchmarks, while requiring only 600MB more memory.
Ternary Bonsai 8B (1.75 GB) reaches 75.5 average benchmark score, compared with 70.5 for 1-bit Bonsai 8B (1.15 GB). Among its peers, it is only behind Qwen3 8B (16.38 GB) and outperforms all other models, despite being 9-10x smaller than them. It posts competitive results across MMLU Redux, MuSR, GSM8K, HumanEval+, IFEval, and BFCLv3, showing that the gain is broad rather than concentrated in a single benchmark.
... continue reading