

Huawei · since November 2026 (Paper arXiv:2611.10649 am 7. November 2026; GitHub-Veröffentlichung um den 10. November 2026) · 2× · last seen Nov 13, 2026
45
Momentum
MACaQ (Model Agnostic Calibrated Quantization) is an open-source quantization tool from Huawei Computing Systems Lab (huawei-csl) for LLMs, MoE models, and diffusion models for image, video, and audio. It quantizes each linear layer independently to 4-bit (MXFP4 W4A4) or produces 2–4-bit GGUF files. It is released under Apache 2.0 on GitHub and integrates with llama.cpp, vLLM, Diffusers, and PyTorch.
Momentum trend
18.08.16.11.
Features
| Throughput/Latency | Only 2.3% end-to-end overhead vs. pure MXFP4 (Qwen3-32B, vLLM) |
| License | Apache License 2.0 |
| Platform | Python/PyTorch; GPU, Ascend CANN kernels via PRs |
| Price | Free, open source (Apache 2.0) |
| Protocol Compatibility | llama.cpp (GGUF), vLLM (W4A4/weight-only), Diffusers, PyTorch |
| Release Date | November 2026 (paper: Nov 7, 2026) |
| Supported Models/Providers | Qwen3 (4B, 27B, 30B, 35B, 80B), OLMoE-1B-7B, FLUX.2-klein-4B, Wan2.2-TI2V, Stable Audio 1.0 |