Home
M6 Intelligence and Tsinghua Open-Source BitCPM-CANN to Run Large Models on Mobile Devices
MBLab, in collaboration with Tsinghua University and the OpenBMB open-source community, has recently released and open-sourced a major breakthrough in low-bit large model training: BitCPM-CANN. Developed natively on Huawei’s Ascend platform, this milestone advances the lightweighting and engineering deployment of edge-side AI large models.

Sixfold Memory Advantage Breaks Hardware Barriers
BitCPM-CANN, now open-source, comes in four sizes: 0.5B, 1B, 3B, and 8B. In direct comparisons with full-precision models of equivalent size, it delivers outstanding performance. Relative to conventional BF16 precision, it achieves approximately six times the memory savings during inference, drastically reducing the hardware needed to run large models.
For mobile devices, this sixfold memory advantage means that 8B-parameter large models—once requiring top-tier hardware—can now run smoothly on mainstream flagship smartphones. This remarkable memory headroom directly accelerates the adoption and commercial deployment of edge-side AI technology on mobile platforms.
High Retention Rate Validates Engineering Reproducibility
Despite the reduction in model size, BitCPM-CANN retains exceptionally high performance, with ability retention rates ranging from 90% to 97.2%. For the three larger model sizes, retention rates stand between 95.7% and 97.2%, while even the smallest 0.5B variant exceeds 90%.
These impressive results systematically demonstrate that the low-bit training approach offers strong scalability and engineering reproducibility. MBLab has established a comprehensive low-bit training foundation using these core technologies, encompassing the full engineering stack—including environment adaptation, 32K long-sequence support, and fused operators—thereby providing a robust public infrastructure for future low-bit training work on the Ascend platform.
Related article
Ando Launches Team Messaging App Bridging Human and AI Collaboration
In 2025, as Sara Du assisted companies in developing MCP servers, a recurring question emerged: how could teams leverage AI agents directly within Slack? The path was obstructed by significant technical barriers, including the complexity of message s
Sequoia-backed Empirik raises $21M to forecast outages before they strike
Before joining Sequoia Capital in 2020 as chief digital and information officer, Avon Puri spent over a decade running infrastructure at Rubrik and VMware. Three years ago, as large language models began showing their true potential, Puri, alongside
OpenAI launches GPT-5.5 Instant as ChatGPT’s new default model
OpenAI launched GPT-5.5 Instant on Monday, replacing GPT-5.3 Instant as the default ChatGPT model. The company states that this update minimizes hallucinations in high-stakes fields like law, medicine, and finance, while preserving the fast response
Related Special Topic Recommendations
Comments (0)
0/500
MBLab, in collaboration with Tsinghua University and the OpenBMB open-source community, has recently released and open-sourced a major breakthrough in low-bit large model training: BitCPM-CANN. Developed natively on Huawei’s Ascend platform, this milestone advances the lightweighting and engineering deployment of edge-side AI large models.

Sixfold Memory Advantage Breaks Hardware Barriers
BitCPM-CANN, now open-source, comes in four sizes: 0.5B, 1B, 3B, and 8B. In direct comparisons with full-precision models of equivalent size, it delivers outstanding performance. Relative to conventional BF16 precision, it achieves approximately six times the memory savings during inference, drastically reducing the hardware needed to run large models.
For mobile devices, this sixfold memory advantage means that 8B-parameter large models—once requiring top-tier hardware—can now run smoothly on mainstream flagship smartphones. This remarkable memory headroom directly accelerates the adoption and commercial deployment of edge-side AI technology on mobile platforms.
High Retention Rate Validates Engineering Reproducibility
Despite the reduction in model size, BitCPM-CANN retains exceptionally high performance, with ability retention rates ranging from 90% to 97.2%. For the three larger model sizes, retention rates stand between 95.7% and 97.2%, while even the smallest 0.5B variant exceeds 90%.
These impressive results systematically demonstrate that the low-bit training approach offers strong scalability and engineering reproducibility. MBLab has established a comprehensive low-bit training foundation using these core technologies, encompassing the full engineering stack—including environment adaptation, 32K long-sequence support, and fused operators—thereby providing a robust public infrastructure for future low-bit training work on the Ascend platform.
Ando Launches Team Messaging App Bridging Human and AI Collaboration
In 2025, as Sara Du assisted companies in developing MCP servers, a recurring question emerged: how could teams leverage AI agents directly within Slack? The path was obstructed by significant technical barriers, including the complexity of message s
Sequoia-backed Empirik raises $21M to forecast outages before they strike
Before joining Sequoia Capital in 2020 as chief digital and information officer, Avon Puri spent over a decade running infrastructure at Rubrik and VMware. Three years ago, as large language models began showing their true potential, Puri, alongside
OpenAI launches GPT-5.5 Instant as ChatGPT’s new default model
OpenAI launched GPT-5.5 Instant on Monday, replacing GPT-5.3 Instant as the default ChatGPT model. The company states that this update minimizes hallucinations in high-stakes fields like law, medicine, and finance, while preserving the fast response











