AI Briefing1.58-bit BitCPM4-CANN released
Key point
OpenBMB released BitCPM4-CANN 1B, 3B, and 8B on Hugging Face.
Details
OpenBMB released the BitCPM4-CANN series on Hugging Face. The linked models are 1B, 3B, and 8B, and the model card lists 4 variants including up to 0.5B. The license is Apache-2.0.
This series is introduced as a family of 1.58-bit ternary models trained on Huawei Ascend NPU. The QAT overhead is around 5%, and on 11 benchmarks, 1B/3B/8B maintained 95.7%~97.2% of the performance of full-precision MiniCPM4. Inference memory was reduced by about 6x.
The models can be loaded like ordinary FP models without any separate quantization library or custom kernels, and ready-to-use examples are provided with Transformers, vLLM, SGLang, and GGUF conversions.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
2026 AI Inference Hardware Innovations: New Architectures for Prefill/Decode Disaggregation and Solving Memory Bottlenecks
Pinterest Reduces Costs by Up to 30% with SSD Serving and Quantization in Manas Embedding Search Platform