Gemma 4 E2B IT — MXFP4 GGUF

Base Model

• Model: google/gemma-4-E2B-it
• Architecture: Gemma4ForConditionalGeneration (Vision + Text)
• Parameters: ~3B (E2B = Efficient 2B-class)
• Context Length: 131,072 tokens (128K)
• License: Apache 2.0
• Update Date: July 20, 2026

Quantization Details

• Format: MXFP4 (OCP Microscaled FP4)
• BPW: 5.66 bits per weight
• File Size: 3.3 GB
• Non-expert tensors: F16

Vision Support

• mmproj: mmproj-gemma-4-E2B-it-f16.gguf (985 MB, F16)
• Supports image and video input

Performance (RTX 5060 Ti 16GB)

• Generation Speed: ~153 t/s at 128K context
• Context: 128K with Q8_0 KV cache fits in 16GB VRAM

Usage (llama.cpp)

llama-cli -m gemma-4-e2b-it-mxfp4.gguf -ngl 99 --flash-attn on -c 131072 --cache-type-k q8_0 --cache-type-v q8_0

Why No MTP?

The Gemma 4 E2B model does not include MTP (Multi-Token Prediction) heads.

Hardware Target

• NVIDIA RTX 50 series (Blackwell) for MXFP4 acceleration
• 16GB+ VRAM recommended