Open Models Almanac
Methodsoptimizationquantization-aware training

general family · filed under optimization

Quantization-Aware Training

Quantization-Aware Training integrates quantization into the training process so the model adapts to the precision loss from quantization.1,2 It simulates quantization during training, which preserves model quality better than post-training quantization when compressing.3

Trade-off Compared with standard post-training quantization, QAT yields higher model quality, enabling smaller memory footprints while preserving capabilities.4

How this definition was written

Summarised by glm-5.3-flash from 2 sources; every sentence is checked against the quote it cites.

  1. QAT integrates the quantization process directly into training — Gemma 4 QAT models: Optimizing model compression for mobile and laptop efficiency
  2. enabling the model, including those of teacher and reference models, to adapt to the precision degradation introduced by quantization — DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence
  3. By simulating quantization during training, QAT minimizes quality loss when the model is compressed — Gemma 4 QAT models: Optimizing model compression for mobile and laptop efficiency
  4. our QAT results yield even higher overall quality compared to standard PTQ baselines — Gemma 4 QAT models: Optimizing model compression for mobile and laptop efficiency

Also called Quantization-Aware Training (QAT).

sources
10
models
6
labs adopt it
5
strongest
core

Documented in

Google DeepMind
Aleph Alpha
Moonshot AI

Evidence

10 quotes from 10 sources, one card per model, each opening with its strongest.

How sources treat it
used 9core 1

Gemma 4 31B

Google DeepMindadopted2 sources

By simulating quantization during training, QAT minimizes quality loss when the model is compressed.

coreoptimizationin Gemma 4 QAT checkpointsGoogle
1 more from this model

We provide quantized versions of our models trained with quantization-aware training [Jacob et al., 2018, QAT] to reduce their parameter memory footprint and latency with minimal impact on quality.

usedpost trainingin Gemma 4Google DeepMind

Kimi K3

Moonshot AIadopted5 sources

Kimi K3 applies quantization-aware training from the SFT stage onward, using MXFP4 weights with MXFP8 activations for broad hardware compatibility.

usedpost trainingin Kimi K3Moonshot AI
4 more from this model

Kimi K3 applies quantization-aware training from the SFT stage onward, using MXFP4 weights with MXFP8 activations for broad hardware compatibility.

usedtraining objectivein Kimi K3Moonshot AI

In addition, we apply quantization-aware training (QAT) from the SFT stage onward, with MXFP4 weights and MXFP8 activations (§ 4.1.4).

usedunclearin Kimi K3Moonshot AI

K3 employs quantization-aware training (QAT) starting from the supervised fine-tuning stage, not post-training quantization.

usedtraining objectivein Kimi K3Moonshot AI

Quantization-aware training starts at the supervised fine-tuning stage, using MXFP4 weights and MXFP8 activations

usedtraining objectivein Kimi K3Moonshot AI
↑