SandLogicTechnologies/gemma-4-E2B-GGUF

gemma-4-E2B

gemma-4-E2B is a multimodal model from the gemma family, designed to handle reasoning tasks across both visual and textual inputs. It is part of a family of models optimized for efficiency, making it suitable for deployment across a wide range of environments including edge devices and local systems.

The model supports multimodal interactions, enabling it to process text and image inputs and generate text outputs. It is optimized for conversational AI, coding, reasoning, and agentic workflows, while maintaining a strong balance between performance and efficiency.

gemma-4-E2B is particularly effective in scenarios involving multi-step reasoning, lightweight deployments, and multilingual communication, while remaining practical for on-device and resource-constrained use cases.

---

Model Overview

• Model Name: gemma-4-E2B
• Architecture: Decoder-only Transformer with multimodal extensions
• Parameter Count: 2B parameters
• Context Window: 128k tokens
• Modalities: Text, Image (multimodal input support)
• Primary Languages: English (with multilingual generalization)
• Developer: Google
• License: Apache 2.0

---

Quantization Details

This repository provides various GGUF quantized versions of the gemma-4-E2B model, optimized for efficient local inference using llama.cpp. Below are the details of the available I-Matrix (IQ) formats.

Quantization Formats (I-Quants)

IQ3_M

• Size reduction of approx 66.33% (2.92 GB) compared to 16-bit (8.67 GB)
• IQ3_M is an aggressive low-bit quantization format designed for scenarios where memory constraints are the primary limitation.
• It compresses model weights down to approximately 3 bits while still attempting to preserve the most important parameters through importance-aware scaling.
• This format is particularly useful when deploying models on highly constrained environments such as edge devices or low-memory systems.
• However, due to the reduced precision, there is a noticeable drop in output quality, especially for reasoning-heavy prompts or long-context tasks, and reconstruction overhead may affect performance depending on the hardware.

IQ4_XS

• Size reduction of approx 64.47% (3.08 GB) compared to 16-bit (8.67 GB)
• IQ4_XS represents a balanced approach between compression and quality.
• It uses a 4-bit representation combined with importance-aware quantization, allowing it to retain significantly more useful information than lower-bit formats while still keeping the model size relatively small.
• This format is often considered a sweet spot for most users, as it delivers strong performance across a wide range of tasks without requiring excessive memory. It performs well in conversational scenarios, general reasoning, and code-related tasks.
• One characteristic of IQ4_XS is that prompt processing may be slightly slower compared to traditional quantization methods due to the added complexity of importance-based reconstruction. However, once generation begins, performance is generally stable and efficient.

IQ4_NL

• Size reduction of approx 63.78% (3.14 GB) compared to 16-bit (8.67 GB)
• IQ4_NL is a more advanced 4-bit quantization format that incorporates non-linear weight mapping in addition to importance-based scaling.
• This allows it to better approximate the original distribution of model weights, particularly in regions where linear quantization would introduce larger errors.
• As a result, IQ4_NL typically provides higher fidelity outputs compared to other 4-bit formats, making it well-suited for tasks that require better reasoning, structured responses, or nuanced language generation.
• The trade-off is a modest increase in model size and computational complexity during dequantization. On some systems, this can lead to slightly slower inference speeds compared to simpler formats, but the quality improvements are often noticeable.

---

Training Overview

Pretraining

The model is trained on a large-scale dataset combining diverse textual corpora and multimodal data sources, enabling it to understand both language and visual information in a unified manner.

Training objectives include:

• Cross-modal representation learning
• Large-scale language modeling
• Visual-text alignment
• Contextual reasoning across long sequences

Alignment and Optimization

Post-training steps refine the model for real-world usability and instruction-following:

• Instruction tuning for conversational tasks
• Reinforcement learning and alignment techniques
• Optimization for reasoning and agentic workflows
• Improved multimodal grounding and response consistency

---

Core Capabilities

• Multimodal understanding Processes text and image inputs to generate context-aware responses.

• Efficient reasoning performance Handles multi-step logical tasks while remaining optimized for smaller-scale deployment.

• Agentic and coding capabilities Supports structured workflows, tool usage patterns, and code-related tasks.

• Long-context processing Maintains coherence across extended inputs and conversations.

• Multilingual support Capable of handling a wide range of languages and mixed-language inputs.

• Deployment flexibility Designed to run effectively across edge devices, local systems, and standard hardware setups.

---

Example Usage

llama.cpp

./llama-cli \
  -m SandlogicTechnologies/gemma-4-E2B_IQ4_NL.gguf /
  -p "Explain the concept of attention in transformer models."

---

Recommended Use Cases

• Lightweight conversational AI systems
• On-device AI applications and edge deployments
• Coding assistants and developer tools
• Multimodal question answering (text + image)
• Educational tools and tutoring systems
• Document analysis and summarization
• Prototyping intelligent applications with limited hardware

---

Acknowledgments

These quantized models are based on the original work by the Google development team.

Special thanks to:

• The Google team for developing and releasing the gemma-4-E2B model.

• Georgi Gerganov and the llama.cpp open-source community for enabling efficient quantization and inference via the GGUF format.

---

Contact

For any inquiries or support, please contact us at support@sandlogic.com or visit our Website.