# huawei-csl/MACaQ

> Calibrated quantization toolkit for text, image, video and audio models down to 2–4 bit, supporting llama.cpp, vLLM, Diffusers and PyTorch. Enables fast compression of large models for local inference.

- Magnitude: 3.5 out of 10 — Cooling
- Stars: 236 total · +4 stars measured 2026-11-14, 00:03–11:14 UTC, ≈ 11 by evening
- Star trust: star growth looks organic
- Category: Language models · Language: C++ · License: Apache-2.0 · Created: 2026-11-10 · Last push: 2026-11-12
- GitHub: https://github.com/huawei-csl/MACaQ · Homepage: https://huawei-csl.com/blog/macaq/ · Page: https://gitnova.dev/en/r/huawei-csl/MACaQ

## Useful for

- Quantize a text LLM to 4-bit GGUF and run it locally with llama.cpp
- Quantize a diffusion model and generate images or video
- Serve a quantized model over an HTTP API with vLLM

## Why it’s here

- Star-counter measurements on 2026-11-14 (UTC), 00:03–11:14: 232 → 236 stars (+4). This is the change over that interval.
- Estimated end-of-day forecast: about +11 stars, using observed gains and the previous day.
- The repository is 4 days old and already has 236 stars. With less than two weeks of history, there's no usual pace to compare the spike against yet.
- Hacker News: “Huawei releases MACaQ, its internal quantizer for local inference” — 134 points, 2 days ago.
- About 8 forks a day — people are taking the code.
- Recent forks include notable developers: @mbrukman (402 followers).

## Star trust

Star growth looks organic.

Star-trust labels are heuristics based on the repository’s behavior, not a check of every stargazer.

## Numbers

- Forks: 13
- Issues and pull requests: 3
- Watchers: 0
- Average over the last week: 49 per day
- Usual pace: too little history (under two weeks)
- Stars in the last hour (measured): 0

## Stars per day, last 5 days (oldest → newest, today is partial)

2026-11-10 … 2026-11-14: 2, 138, 79, 17, 4

## Magnitude by day, last 4 days

11-11 3.0, 11-12 6.6, 11-13 5.9, 11-14 3.8

## Hacker News

- Huawei releases MACaQ, its internal quantizer for local inference — 134 points, 32 comments: https://news.ycombinator.com/item?id=50458166

## Spotted in now

- Spotted on Hacker News

## Similar by description

1. **NVIDIA/Model-Optimizer** — 1.1 · Steady · Language models · Python · +2 stars measured 2026-11-14, 00:12–11:20 UTC, ≈ 5 by evening
   NVIDIA library for compressing and accelerating models: quantization, pruning, distillation, NAS and speculative decoding with export to TensorRT-LLM, vLLM, SGLang. For ML engineers preparing models for deployment.
   Full card: https://gitnova.dev/en/r/NVIDIA/Model-Optimizer.md
2. **techjarves/Portable-Local-Studio** — 1.1 · Cooling · Apps · JavaScript · +4 stars measured 2026-11-14, 00:11–11:20 UTC, ≈ 8 by evening
   Offline local AI studio for Windows, Linux and macOS with a zero-setup GUI: image generation (Stable Diffusion), GGUF LLM chat, speech-to-text (Whisper) and text-to-speech (Kokoro TTS).
   Full card: https://gitnova.dev/en/r/techjarves/Portable-Local-Studio.md
3. **leejet/stable-diffusion.cpp** — 1.5 · Steady · Generative media · C++ · +3 stars measured 2026-11-14, 00:09–11:19 UTC, ≈ 8 by evening
   Diffusion model inference (Stable Diffusion, FLUX, Wan, Qwen Image, etc.) implemented in pure C/C++ on top of ggml — like llama.cpp but for image and video generation. Runs without external dependencies on CPU, CUDA, Vulkan, Metal and…
   Full card: https://gitnova.dev/en/r/leejet/stable-diffusion.cpp.md
4. **JakeATX/llamAmpere** — 0.9 · Steady · Language models · C++ · +0 stars measured 2026-11-14, 00:07–11:20 UTC, ≈ 1 by evening
   A llama.cpp fork tuned for NVIDIA Ampere cards (RTX 3090/3090 Ti) to run large Qwen3.8 models fast with long context and partial expert offload to system RAM.
   Full card: https://gitnova.dev/en/r/JakeATX/llamAmpere.md

---

Magnitude (0–10) measures how fast and how unusually interest in a repository is growing right now. It is not a quality score.
Days are UTC. “So far today” is a fact; “expected by the end of the day” is a forecast.
Summaries and use cases are written by an LLM (DeepSeek V4.1 Flash) from the README and may be inaccurate: verify specific claims (benchmarks, speed, hardware) in the repository itself.
Data as of 2026-11-14 11:22 UTC, updated every 30 minutes.