Skip to content
Computer vision

Gemma 4 E4B

compact multimodal model from Google

2 min readupdated

Gemma 4 E4B is an open multimodal model from Google DeepMind that takes text, images, video, and audio as input and generates text output. While it has 8B total parameters, PLE (Per-Layer Embeddings — separate embedding tables for each layer) brings the active count down to just 4.5B effective parameters, making it lightweight enough to run on laptops and mobile devices. Unlike its larger siblings (26B and 31B), E4B packs a built-in audio encoder, native thinking mode, and system prompt support. It's a solid fit for developers building local assistants, agents, and apps that need to process multiple data types offline without relying on cloud APIs.

Key features

  • Full multimodality — handles text, images at any aspect ratio, video (frame sequences), and audio (speech recognition and translation).
  • Efficient architecture — 42 layers with hybrid attention (alternating 512-token local windows and global attention) and 4.5B active parameters out of 8B total.
  • Thinking mode and function calling — reasons step-by-step before answering (enable_thinking=True), supports the system role, and offers native function calling for agent workflows.
  • Long context and multilingual support — up to a 128K token context window, pretrained on 140+ languages with out-of-the-box support for 35+ languages.
  • Permissive license — released under the Apache 2.0 license in both base and instruction-tuned (-it) variants.

What you can do

  • Local voice assistant — use the built-in audio encoder for on-device automatic speech recognition (ASR) and speech-to-text translation.
  • Document and screen parsing — extract handwritten text, charts, PDFs, and UI elements inside local automation scripts.
  • Autonomous tool-using agent — wire up the model via transformers with native function calling and system prompts to execute multi-step tasks.

Example

from transformers import AutoProcessor, AutoModelForMultimodalLM

MODEL_ID = "google/gemma-4-E4B-it"
processor = AutoProcessor.from_pretrained(MODEL_ID)
model = AutoModelForMultimodalLM.from_pretrained(MODEL_ID, dtype="auto", device_map="auto")

messages = [{"role": "user", "content": "Write a short joke about saving RAM."}]
inputs = processor.apply_chat_template(messages, tokenize=True, return_dict=True, return_tensors="pt", add_generation_prompt=True).to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256)
Read moreDetails
Benchmarks (-it version): MMLU Pro — 69.4%, MMMLU — 76.6%, LiveCodeBench v6 — 52.0%, MATH-Vision — 59.5%, FLEURS (audio, lower is better) — 0.08.
Architecture: 8B total parameters (4.5B effective), ~150M in the vision encoder and ~300M in the audio encoder, 262K token vocabulary.
Optimized for on-device inference on laptops and smartphones, supported in the latest version of transformers.
Discuss Telegram 0 22

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 15 GB · ready files17 GB
  • FP8 / INT8recommendedweights 7.4 GB8.7 GB
  • 4-bitweights 4.2 GB5.1 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 3080 · BF16 — from RTX 4090

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # mdcreativeagent/gemma-4-E4B
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download mdcreativeagent/gemma-4-E4B --local-dir ~/models/gemma-4-E4B
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to gemma-4-E4B
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of gemma-4-E4B (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

mdcreativeagent/gemma-4-E4B
Author
mdcreativeagent
Category
Computer vision
Released
Posted

On NeyroLearn

Hugging Face

220

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments