Open Models Almanac
Methodsinference & servinginference quantization

specific method · filed under inference & serving

Mobile quantization

Mobile quantization is an inference-quantization schema engineered for mobile and edge hardware.1 It pre-calculates activation scaling during training, quantizes channel-wise to match mobile accelerators, and compresses token-generating parts more aggressively while keeping core layers at higher precision.2

Trade-off It saves storage and memory without degrading the model's quality.3

How this definition was written

Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.

  1. we engineered a custom mobile-quantization schema designed for edge hardware — Gemma 4 QAT models: Optimizing model compression for mobile and laptop efficiency
  2. Static activations: Normally, models waste processing power calculating how to scale data on the fly. We pre-calculate these settings during training — Gemma 4 QAT models: Optimizing model compression for mobile and laptop efficiency
  3. This saves storage without making the model less smart — Gemma 4 QAT models: Optimizing model compression for mobile and laptop efficiency

Also called mobile-specialized quantization schema, custom mobile-quantization schema.

sources
2
model
1
lab adopt it
1
strongest
core

Documented in

Google DeepMind

Evidence

2 quotes from 2 sources, one card per model, each opening with its strongest.

How sources treat it
optional 1core 1

Gemma 4 31B

Google DeepMindadopted2 sources

we engineered a custom mobile-quantization schema designed for edge hardware

coreoptimizationin Gemma 4 E2B and E4B edge modelsGoogle
1 more from this model

mobile quantization: per-channel low bitwidth weight (mix of int2 and int4) and activation quantization (int8).

optionalinference servingin Gemma 4Google DeepMind
↑