BitCPM-CANN has been launched as the world’s first open-sourced 1.58-bit ternary LLM, developed by ModelBest, Tsinghua University, and the OpenBMB community, utilizing Chinese-developed AI infrastructure on Huawei Ascend 910B NPUs. This model utilizes only three weight states, allowing for significantly reduced memory requirements when deployed on various devices such as phones, PCs, and cars, which is crucial given the current trend of skyrocketing hardware costs. Furthermore, the entire training process was executed natively on the Ascend architecture, ensuring high performance and compatibility, and the project emphasizes open-source reproducibility by releasing complete training scripts and verification methods for community use.
Huawei: Huawei develops advanced AI hardware including the Ascend series of NPUs designed for high-performance computing and inference. Its Ascend 910B platform served as the native environment for the entire end-to-end training pipeline from operators to frameworks. This integration demonstrates support for specialized, China-developed infrastructure in large-scale model development.
OpenBMB: OpenBMB is an open lab dedicated to building foundation models and systems aimed at advancing toward AGI, with a focus on efficient and accessible AI technologies. It coordinated the community release and verification of the full model family, ensuring reproducibility across benchmarks. The initiative underscores its role in open-sourcing practical innovations for edge AI.
ModelBest: ModelBest is a Chinese AI company specializing in on-device and edge AI solutions, including intelligent agents and efficient model architectures. It led key aspects of the quantization and training innovations for this project while partnering with academic and open-source groups. The work aligns with its ongoing focus on practical, deployable AI systems for mobile, PC, and automotive environments.
BitCPM-CANN: BitCPM-CANN is an open-source 1.58-bit ternary large language model developed through a collaboration focused on efficient AI training and deployment. It represents a full-pipeline quantized model family spanning multiple sizes, built with quantization operators, algorithms, and a complete training framework. The project directly addresses rising hardware costs and edge deployment needs by enabling the model to run on resource-constrained devices without new hardware.
Tsinghua University: Tsinghua University is a leading Chinese research institution with strong programs in computer science and artificial intelligence. Its researchers contributed to the algorithmic and framework development for the ternary quantization pipeline. The collaboration highlights academic efforts to advance reproducible, hardware-native AI training methods.
`json
{
“Edge Deployment Focus”: “The approach is designed to enhance memory efficiency, enabling deployment of more advanced models on personal and industrial devices without extra hardware requirements.”,
“Hardware-Native Training”: “The training framework for low-bit models was completely developed and executed on Huawei Ascend NPUs, ensuring seamless integration with the hardware.”,
“Open-Source Reproducibility”: “Comprehensive training scripts and model verifications have been made available to facilitate research, deployment, and ongoing community improvements.”
}
`