Produced with the codec from "Balanced Ternary Post-Training Quantization for Large Language Models" (Stentzel, 2026). See Entrit/tritllm-codec for the codec source.

Quick load

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("Entrit/Qwen2.5-14B-trit-uniform-d4")
tokenizer = AutoTokenizer.from_pretrained("Entrit/Qwen2.5-14B-trit-uniform-d4")

The weights are dequantized to FP16 for stock-transformers compatibility. The on-disk size is therefore the same as the FP16 source. The 6.64-bpw figure refers to the information content of the quantized matrices and is what matters for inference on hardware that consumes the packed trit format directly (see Entrit/tritllm-kernel).

Quantization details

Field	Value
Source model	`Qwen/Qwen2.5-14B`
Depth	d=4 (81 levels)
Bits per weight	6.64
Group size	16
Scale codebook	27-entry log-spaced (scale_depth=3)
Method	Uniform PTQ
Quantized layers	all 2D linear matrices
Kept FP16	`lm_head`, token embeddings, all `*_norm` layers
Codec	tritllm v2

Citation

@article{stentzel2026ternaryptq,
  title  = {Balanced Ternary Post-Training Quantization for Large Language Models},
  author = {Stentzel, Eric},
  year   = 2026,
  note   = {Entrit Systems}
}

Reproducibility

bash

git clone https://huggingface.co/Entrit/tritllm-codec
cd tritllm-codec
python quantize_model_v2.py --model Qwen/Qwen2.5-14B --configs uniform-d4 --out ./out