Home
Mianbi Intelligence and Tsinghua University Unveil China’s First 1.58-bit Large Model BitCPM-CANN
Mingshi Intelligence, in partnership with Tsinghua University and the OpenBMB community, has officially open-sourced BitCPM-CANN, China’s first large language model trained on Huawei’s Ascend platform. Utilizing ternary (1.58-bit) quantization, this breakthrough advances low-bit training technology and marks a significant milestone for China’s AI ecosystem.
BitCPM-CANN showcases the robust capabilities of domestic computing infrastructure, establishing a fully native development pipeline from quantization operators to training algorithms. Available in 0.5B, 1B, 3B, and 8B variants, it outperforms full-precision MiniCPM4 models of equivalent size. Notably, it reduces inference memory usage by approximately sixfold, enabling smooth operation of 8B parameter models on mainstream flagship smartphones and offering substantial benefits to the mobile industry.

Powered by MindSpeed and Megatron-LM, Mingshi Intelligence has constructed a comprehensive low-bit training framework. This infrastructure supports environment adaptation, 32K long-sequence handling, parallel strategies, and integrated operators. By providing this public foundation for all Ascend-based low-bit training, the company lowers development barriers and accelerates technological iteration.

To foster broader adoption, all BitCPM-CANN model weights are open-sourced and accessible via HuggingFace and ModelScope. This initiative equips developers with powerful tools, encouraging innovative applications across the AI landscape.
In conclusion, BitCPM-CANN represents a crucial advancement in China’s AI large model training capabilities, laying a strong foundation for future intelligent applications.
Related article
Meta Joins TSMC and Micron to Equip New Servers with Old Memory
As artificial intelligence adoption accelerates worldwide, next-generation DDR5 memory is experiencing a critical supply shortage. To avoid delays in AI data center deployments, Meta has partnered with industry leaders like TSMC and Micron to enginee
How Time Series Databases Power Real-Time Robotics Data
Doug Pagnutti of Tiger Data.Pagnutti serves as an industrial developer advocate at Tiger Data, supporting teams in building real-world industrial systems. With over 12 years of experience in oil and gas and manufacturing automation, he worked on PLCs
How to optimize SEO for Google Japan and Yahoo Japan?
Master Loom: Create and Share Professional Videos with AITable of ContentsIntroduction to LoomAccessing LoomRecording OptionsFull ScreenWindowCameraAudio and Sound QualityEditing and CustomizationBackground OptionsEditing the VideoAdding EffectsShari
Related Special Topic Recommendations
Comments (0)
0/500
Mingshi Intelligence, in partnership with Tsinghua University and the OpenBMB community, has officially open-sourced BitCPM-CANN, China’s first large language model trained on Huawei’s Ascend platform. Utilizing ternary (1.58-bit) quantization, this breakthrough advances low-bit training technology and marks a significant milestone for China’s AI ecosystem.
BitCPM-CANN showcases the robust capabilities of domestic computing infrastructure, establishing a fully native development pipeline from quantization operators to training algorithms. Available in 0.5B, 1B, 3B, and 8B variants, it outperforms full-precision MiniCPM4 models of equivalent size. Notably, it reduces inference memory usage by approximately sixfold, enabling smooth operation of 8B parameter models on mainstream flagship smartphones and offering substantial benefits to the mobile industry.

Powered by MindSpeed and Megatron-LM, Mingshi Intelligence has constructed a comprehensive low-bit training framework. This infrastructure supports environment adaptation, 32K long-sequence handling, parallel strategies, and integrated operators. By providing this public foundation for all Ascend-based low-bit training, the company lowers development barriers and accelerates technological iteration.

To foster broader adoption, all BitCPM-CANN model weights are open-sourced and accessible via HuggingFace and ModelScope. This initiative equips developers with powerful tools, encouraging innovative applications across the AI landscape.
In conclusion, BitCPM-CANN represents a crucial advancement in China’s AI large model training capabilities, laying a strong foundation for future intelligent applications.
Meta Joins TSMC and Micron to Equip New Servers with Old Memory
As artificial intelligence adoption accelerates worldwide, next-generation DDR5 memory is experiencing a critical supply shortage. To avoid delays in AI data center deployments, Meta has partnered with industry leaders like TSMC and Micron to enginee
How Time Series Databases Power Real-Time Robotics Data
Doug Pagnutti of Tiger Data.Pagnutti serves as an industrial developer advocate at Tiger Data, supporting teams in building real-world industrial systems. With over 12 years of experience in oil and gas and manufacturing automation, he worked on PLCs
How to optimize SEO for Google Japan and Yahoo Japan?
Master Loom: Create and Share Professional Videos with AITable of ContentsIntroduction to LoomAccessing LoomRecording OptionsFull ScreenWindowCameraAudio and Sound QualityEditing and CustomizationBackground OptionsEditing the VideoAdding EffectsShari











