- Who it helps
- Researchers and inference engineers evaluating model compression
- Availability
- November 7 paper and Apache-2.0 implementation are public. Repository folders cover several inference integrations; their compatibility was not tested.
What calibration tries to preserve
Quantization uses fewer bits to represent numbers. MACaQ adjusts this representation using sample activations, while Flip-Refinement changes selected numeric codes to reduce conversion error.
Separate conversion from serving
The reported conversion-memory result concerns preparing a compressed model. It does not establish that the resulting model fits or runs acceptably on every consumer GPU. Quality and latency still depend on the model, data and runtime.
What you can do
- Read the primary description and its stated limits before planning an evaluation. Open reference ↗
Where the evidence stops
Paper abstract and implementation README checked November 12. No model converted, loaded or benchmarked.
Sources read. Public source text checked November 12, 2026 UTC. No third-party project or full workflow executed.
Sources & further reading
Paper · paper
MACaQ paper ↗
November 7 abstract read; conversion-memory and benchmark figures are author reports.
Lorenz Muller, Philippe Bich and coauthors · page checked 2026-11-12 · Published 2026-11-07
GitHub · implementation
MACaQ implementation ↗
Apache-2.0 repository and supported integration folders read. No compatibility test executed.
Huawei CSL · page checked 2026-11-12
82 stars · observed 2026-11-12 00:34 UTC
Hugging Face · discussion
MACaQ paper entry ↗
Author-associated paper entry; not an independent reproduction.
pbicho · page checked 2026-11-12
Coverage note: Paper abstract and implementation README checked November 12. No model converted, loaded or benchmarked.
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
- · Added to Current
MACaQ converts model weights and intermediate values to fewer-bit representations. It calibrates one layer at a time to limit the memory needed during conversion.