SandLogicTechnologies/gemma-4-E4B-GGUF

gemma-4-E4B

gemma-4-E4B is a multimodal model from the gemma family, built to handle more demanding reasoning and generation tasks across both visual and textual inputs. With a larger parameter count, it offers improved reasoning depth, stronger consistency across long contexts, and better performance on complex problem-solving workloads.

The model supports multimodal interactions, allowing it to process text, images, and structured content such as long-form documents. It is designed for advanced conversational systems, agentic pipelines, and applications that require higher accuracy and structured outputs.

gemma-4-E4B is particularly suited for tasks that involve multi-step reasoning, technical workflows, and multilingual processing, while still maintaining practical efficiency for optimized deployments.

---

Model Overview

• Model Name: gemma-4-E4B
• Architecture: Decoder-only Transformer with multimodal extensions
• Parameter Count: 4B 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-E4B 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 (3-bit Medium)

• IQ3_M prioritizes extreme compression, enabling the model to run within very tight memory constraints while still preserving essential behavior.
• It represents weights at approximately 3 bits per parameter and uses importance-aware scaling to retain the most impactful components.
• This format is useful for experimental setups, edge environments, or scenarios where fitting the model is the primary constraint.
• Due to the aggressive reduction in precision, performance on complex reasoning, long-context tasks, and structured outputs can degrade, and reconstruction overhead may influence runtime efficiency.
• Size reduction of approx 68.7% (4.39 GB) compared to 16-bit (14.02 GB)

IQ4_XS

• IQ4_XS provides a compact yet capable configuration by combining 4-bit quantization with importance-driven weighting.
• It maintains a practical balance between size and performance, making it suitable for a wide range of real-world applications.
• This format handles conversational tasks, coding prompts, and moderate reasoning workloads effectively without requiring large memory budgets.
• While internally more complex than traditional quantization, it delivers consistent generation performance once inference begins.
• Size reduction of approx 66.2% (4.74 GB) compared to 16-bit (14.02 GB)

IQ4_NL

• IQ4_NL introduces non-linear transformations to the quantization process, allowing it to better model variations in weight distributions.
• This results in improved fidelity for critical layers, particularly in scenarios involving structured reasoning and long-form generation.
• It is well-suited for higher-precision workloads such as technical explanations, analytical outputs, and agentic task execution.
• The increased representational quality comes with slightly higher computational cost and marginally larger model size compared to simpler formats.
• Size reduction of approx 65.4% (4.85 GB) compared to 16-bit (14.02 GB)

---

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-heavy and agentic workflows
• Enhanced multimodal grounding and response consistency

---

Core Capabilities

• Deep multi-step reasoning Handles complex chains of thought and maintains logical consistency across extended problem-solving tasks.

• Agentic workflow support Enables structured task execution, tool-like reasoning patterns, and multi-stage interactions.

• High-fidelity multimodal understanding Combines visual and textual signals to produce context-aware and precise outputs.

• Advanced technical generation Produces detailed, structured, and accurate responses for coding, analysis, and technical domains.

• Long-context stability Maintains coherence and relevance across extended documents and multi-turn conversations.

• Multilingual robustness Performs reliably across diverse languages and mixed-language inputs.

---

Example Usage

llama.cpp

./llama-cli \
  -m SandlogicTechnologies/gemma-4-E4B_IQ4_NL.gguf / \
  -p "Explain the concept of attention in transformer models."

---

Recommended Use Cases

• Advanced reasoning systems and analytical pipelines
• Agentic AI workflows and structured task automation
• Technical research assistants and domain-specific copilots
• Multimodal document and data interpretation
• Code generation, debugging, and system design
• Long-context knowledge extraction and summarization
• Educational platforms for complex problem solving
• High-quality conversational agents requiring consistency and depth

---

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-E4B 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.