Models

The PAUL Open model registry. All models are fine-tuned from the Google Gemma 4 family and are designed for consumer hardware reproducibility.

paul-open-e4b-it

Built

Primary experimental pilot. Fits easily on a single T4 or RTX 3060.

Parameters
4.5B
VRAM (4-bit)
~6GB
Context
8K
Architecture
Gemma 4

paul-open-12b-it

Planned

Intermediate target. Requires an RTX 3090/4090 for 4-bit inference.

Parameters
12B
VRAM (4-bit)
~12GB
Context
8K
Architecture
Gemma 4

paul-open-26b-it

Planned

Full capability target. Requires 24GB VRAM for 4-bit, or dual cards for unquantized inference.

Parameters
26B
VRAM (4-bit)
~24GB
Context
8K
Architecture
Gemma 4

paul-open-31b-it

Planned

Flagship model. Requires dual consumer GPUs or professional hardware.

Parameters
31B
VRAM (4-bit)
~28GB
Context
8K
Architecture
Gemma 4

Hardware Requirements

TierParametersContextMin VRAM (Inference)Target Hardware
e2b-it2.6B8K~4GB (4-bit)Laptops / RTX 3050
e4b-it4.5B8K~6GB (4-bit)RTX 3060 / T4 (Primary Pilot)
12b-it12B8K~12GB (4-bit)RTX 3090 / 4090
26b-it26B8K~24GB (4-bit)RTX 3090 / 4090 / A100