No official torrent yet β requesting it tells the flock what to host next. You cannot seed a catalog-only page.
Cataloged from openbmb/BitCPM-CANN-0.5B-unquantized
All models by openbmb β
GitHub Repo | Technical Report
π Join us on Discord and WeChat
BitCPM-CANN-0.5B-unquantized is the unquantized QAT (Quantization-Aware Training) checkpoint of BitCPM-CANN-0.5B, designed for continued pre-training and fine-tuning. It preserves full-precision latent weights with ternary fake quantizers (weights β {-1, 0, 1} with group-wise scaling, trained via STE) defined in modeling.py, enabling the model to keep learning under quantization constraints. For technical details, see our Technical Report.
β οΈ This model is NOT for direct inference. For inference, use the pseudo-quantized version: openbmb/BitCPM-CANN-0.5B.
The only requirement is that the forward pass must go through the bundled modeling.py (which contains the ternary fake quantizer). Load with trust_remote_code=True and do NOT replace or bypass the model's forward logic.
We provide ready-to-use training scripts in the example directory (using the 1B model as an example):
example/run.sh + example/train.pyexample/run_sft.sh + example/train_sft.pyQuick start:
Any framework that supports HuggingFace model loading with custom code can be used, such as LLaMA Factory, HuggingFace Trainer, etc. The key is to ensure trust_remote_code=True:
After training, use qat-convert.py to fuse the fake quantizer and produce inference-ready pseudo-quantized weights:
The converted model can be loaded for inference in the same way as openbmb/BitCPM-CANN-0.5Bβno special quantization libraries required.
# Continued pre-trainingcd example && bash run.sh
# Supervised fine-tuningcd example && bash run_sft.shfrom transformers import AutoModelForCausalLM, AutoTokenizer
path = 'openbmb/BitCPM-CANN-0.5B-unquantized'tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True)model = AutoModelForCausalLM.from_pretrained( path, torch_dtype=torch.bfloat16, trust_remote_code=True)
# Use with your preferred framework (LLaMA Factory, HF Trainer, etc.)# The ternary fake quantizer in modeling.py is applied automatically during forward pass.python qat-convert.py \ --input_bin <path-to-finetuned-pytorch.bin> \ --output <path-to-output-pseudo-quantized-pytorch.bin> \ --quant_type ternary \ --group_size -1ββββββββββββββββββββββββββββββββββββ BitCPM-CANN-0.5B-unquantized β β This model (QAT checkpoint + fake quantizer in modeling.py)βββββββββββββββββ¬ββββββββββββββββββ β βΌ Train (DeepSpeed / LLaMA Factory / HF Trainer / ...)ββββββββββββββββββββββββββββββββββββ Fine-tuned checkpoint β β Still contains un-fused QAT parametersβββββββββββββββββ¬ββββββββββββββββββ β βΌ python qat-convert.py --quant_type ternary --group_size -1ββββββββββββββββββββββββββββββββββββ Pseudo-quantized model β β Ready for inference (same format as BitCPM-CANN-0.5B)βββββββββββββββββββββββββββββββββββ@article{bitcpmcann, title={{BitCPM-CANN}: Native 1.58-Bit Large Language Model Training on Ascend NPU}, author={BitCPM Team}, year={2026}}