GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
Entrit
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
Entrit
Model Tree
Qwen/Qwen2.5-72B
Input Modalities
Output Modalities
Supported Functionality
python
from transformers import AutoModelForCausalLM, AutoTokenizermodel = AutoModelForCausalLM.from_pretrained("Entrit/Qwen2.5-72B-trit-uniform-d4")tokenizer = AutoTokenizer.from_pretrained("Entrit/Qwen2.5-72B-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).
| Field | Value |
|---|---|
| Source model | Qwen/Qwen2.5-72B |
| 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 |
Following standard quantization-paper convention, only the 2D linear weight matrices are ternary-quantized. The token embedding lookup and final classifier (lm_head) stay in FP16. The 6.64-bpw figure is computed over the quantized matrices only, consistent with how GPTQ, AWQ, and NF4 report BPW.
markdown
@article{stentzel2026ternaryptq,title = {Balanced Ternary Post-Training Quantization for Large Language Models},author = {Stentzel, Eric},year = 2026,note = {Entrit Systems}}
bash
git clone https://huggingface.co/Entrit/tritllm-codeccd tritllm-codecpython quantize_model_v2.py --model Qwen/Qwen2.5-72B --configs uniform-d4 --out ./out
The output at ./out/uniform-d4/model/ matches this repository.