Gemma 3 12B IT - NVFP4 Quantized

This is a FP4 (4-bit floating point) quantized version of google/gemma-3-12b-it, optimized for NVIDIA GPUs with native FP4 support (Blackwell architecture and newer).

Model Details

Attribute	Value
Base Model	google/gemma-3-12b-it
Quantization	NVFP4 (NVIDIA 4-bit floating point)
Target Hardware	NVIDIA Blackwell GPUs (B100, B200, GB200)
Original Parameters	12B

Description

This model provides pre-quantized NVFP4 weights for Gemma 3 12B Instruct, enabling efficient inference on NVIDIA's Blackwell architecture GPUs with native FP4 tensor core support. Loading pre-quantized weights avoids the overhead of runtime quantization.

Why NVFP4?

Native hardware support: Blackwell GPUs include dedicated FP4 tensor cores
~4x memory reduction: Compared to FP16/BF16 weights
Faster inference: Leverages hardware-accelerated FP4 matrix operations
Pre-quantized: No quantization overhead at load time

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "yepthatsjason/gemma-3-12b-it-nvfp4"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
)

Hardware Requirements

Required: NVIDIA Blackwell GPU (B100, B200, GB200, or newer with FP4 support)
VRAM: ~6GB (significantly reduced from ~24GB for BF16)

Quantization Details

This model was quantized using NVIDIA's FP4 format, which uses 4 bits per weight with a floating-point representation optimized for neural network inference on Blackwell architecture.

Limitations

Requires Blackwell or newer NVIDIA GPUs with native FP4 support
May show slight accuracy degradation compared to full-precision model
Not compatible with older GPU architectures (Ampere, Hopper without FP4 emulation)