🧠

Multimodal Language Models – Gemma 4 E2B

🛠

On-Device Text, Image & Audio Understanding

Google DeepMind's smallest and most efficient Gemma 4 model, designed to run locally on phones and laptops, with full support for text, images, and audio.

What it does

Gemma 4 E2B is the smallest model in Google DeepMind's Gemma 4 family, built specifically for on-device and edge deployment. Despite its compact size, it is a fully multimodal AI that understands text, images, and audio, and is one of only two models in the Gemma 4 family to natively support audio input. It handles a wide range of everyday AI tasks including answering questions, analyzing images, transcribing and translating speech, summarizing documents, generating code, and powering conversational assistants. The "E" in E2B stands for "effective" and the model is architecturally optimized to deliver strong performance well above what its size would normally suggest.

Problem it solves

  • On-device AI – Runs efficiently on smartphones and laptops without requiring cloud connectivity, enabling private, low-latency AI experiences on consumer hardware
  • Audio understanding – One of only two Gemma 4 models to natively support audio input, enabling speech transcription and translation directly within the model — no separate speech processing pipeline needed
  • Multimodal in a small package – Handles text, images, and audio together in a single model at a size previously unavailable in open-weight models of this quality
  • Cost-efficient deployment – The compact size makes it dramatically cheaper to run at scale in cloud environments compared to larger models, without sacrificing the ability to handle real-world multimodal tasks
  • Privacy-friendly – Designed for local execution, making it well-suited for applications where data should not leave the device

Input/Output

  • Input:
    • Text – Questions, instructions, documents, code, and conversations
    • Images – Photos, screenshots, charts, diagrams, and documents at variable resolution and aspect ratio
    • Audio – Up to 30 seconds of speech for transcription or translation
    • Video – Up to 60 seconds of video processed as sequential frames (one frame per second)
    • Interleaved – Text, images, and audio can be freely mixed in any order within a single prompt
  • Advanced:
    • Max New Tokens: Controls the maximum length of the model's response. Higher values allow longer, more detailed outputs; lower values keep responses concise. Default recommendation is 1,024 tokens for most tasks.
    • Temperature: Controls how creative or predictable the response is. Set to 0 for fully deterministic, consistent outputs (best for factual or structured tasks). Higher values produce more varied and creative responses. Google recommends 1.0 for general use.
    • Top-P (Nucleus Sampling): Sets a probability threshold limiting which words the model considers. A Top-P of 0.95 means the model only picks from the most probable words that together add up to 95% — filtering out unlikely choices. Lower values make responses more focused; higher values allow more variety. Recommended: 0.95.
    • Top-K Sampling: Limits the model to choosing from only the top K most likely next words. A Top-K of 64 means the model considers only the 64 most probable options at each step. Works alongside Top-P to keep responses coherent. Recommended: 64.
    • Enable Thinking Mode: When turned on, the model reasons step-by-step before delivering its final answer — similar to showing its work. Significantly improves accuracy on complex tasks. When turned off, the model responds directly and more quickly — better for simple or conversational tasks.
  • Output:
    • Generated text responses, analysis, transcriptions, translations, summaries, or code
    • Supports a built-in thinking mode where the model reasons step-by-step before delivering its final answer

Accuracy & Speed

Benchmark Performance

Benchmark
Gemma 4 E2B
What it measures
MMLU-Pro
60.0%
Broad expert knowledge
AIME 2026
37.5%
Advanced mathematics
LiveCodeBench v6
44.0%
Real-world coding ability
GPQA Diamond
43.4%
Expert-level science reasoning
MMMLU (multilingual)
67.4%
Cross-language knowledge
MMMU-Pro
44.2%
Multimodal academic understanding
FLEURS (speech, lower is better)
0.09
Speech recognition accuracy
For context: these scores significantly outperform the previous generation Gemma 3 27B model on reasoning and coding, despite being a fraction of the size.

Technical Details

Architecture
Dense transformer with hybrid local + global attention
Effective parameters
2.3 billion (5.1B including embeddings)
Context window
128,000 tokens (~100,000 words)
Modalities supported
Text, Image, Audio, Video
Audio support
Speech recognition (ASR) and speech translation (AST) — up to 30 seconds
Video support
Up to 60 seconds at 1 frame per second
Thinking mode
Built-in — enable for step-by-step reasoning on complex tasks
Function calling
Native — connects to external tools and APIs for agentic workflows
Languages
35+ out-of-the-box; pre-trained on 140+ languages
Supported frameworks
Hugging Face Transformers

Compliance & Provenance

Provider
Google (open weights)
Provider type
GPAI — General-Purpose AI (GPAI)
License
Apache 2.0
EU AI Act risk class
Limited Risk
Art. 50 transparency
Required — outputs are marked. See AI Policy §2.
Region availability
Available globally
Training data summary
Provider model card / training data

For more on how we classify models and mark outputs, see our AI Policy.

Model Source