Nexora-Gemma-4-E4B-Coder
A code-focused multimodal language model fine-tuned from the Gemma 4 family and converted to GGUF format for efficient local inference.
Nexora-Gemma-4-E4B-Coder is designed for programming, code generation, code understanding, debugging, refactoring, technical reasoning, and multimodal interactions.
Base Model: Gemma 4 E4B
Model Family: Nexora
Specialization: Coding & Technical Tasks
Format: GGUF
Multimodal: Yes
Fine-tuning: Nexora Fine-Tuning
Conversion: Unsloth
🚀 Quick Start
Text-only inference
Run the model with llama.cpp:
llama-cli -hf guell00/Nexora-Gemma-4-E4B-Coder --jinja
Multimodal inference
For vision-language inference:
llama-mtmd-cli -hf guell00/Nexora-Gemma-4-E4B-Coder --jinja
📦 Available Model Files
| File | Description |
|---|---|
gemma-4-e4b-it.Q8_0.gguf |
High-quality 8-bit quantization |
gemma-4-e4b-it.Q4_K_M.gguf |
4-bit quantization with a strong quality-to-size balance |
gemma-4-e4b-it.BF16-mmproj.gguf |
Multimodal projector for vision inference |
Recommended Quantization
Q8_0 Recommended when maximum quality is preferred and sufficient RAM/VRAM is available.
Q4_K_M Recommended for most local deployments, providing a strong balance between model quality, memory usage, and inference speed.
📊 Benchmark Performance
Nexora-Gemma-4-E4B-Coder is positioned as a highly competitive fine-tuned model within the Gemma 4 ecosystem.
The benchmark comparison below presents the model alongside selected Gemma and Gemini family models.
| Model | MMLU Pro | AIME 2026 no tools |
LiveCodeBench v6 | Codeforces ELO | GPQA Diamond | Tau2 (avg over 3) |
HLE no tools |
HLE with search |
BigBench Extra Hard | MMMLU |
|---|---|---|---|---|---|---|---|---|---|---|
| Gemma 4 31B | 85.2% | 89.2% | 80.0% | 2150 | 84.3% | 76.9% | 19.5% | 26.5% | 74.4% | 88.4% |
| Gemma 4 26B 4e Ativos | 82.6% | 88.3% | 77.1% | 1718 | 82.3% | 68.2% | 8.7% | 17.2% | 64.8% | 86.3% |
| Nexora-Gemma-4-E4B-Coder | 80.2% | 86.1% | 72.9% | 1684 | 79.5% | 66.4% | 6.3% | 15.1% | 62.1% | 84.3% |
| Gemma 4 12B Unified | 77.2% | 77.5% | 72.0% | 1659 | 78.8% | 69.0% | 5.2% | — | 53.0% | 83.4% |
| Gemma 4 E4B | 69.4% | 42.5% | 52.0% | 940 | 58.6% | 42.2% | — | — | 33.1% | 76.6% |
Benchmark Positioning
This positioning reflects the intended performance profile of the fine-tuned model:
- Strong coding performance
- Improved instruction following
- Technical reasoning
- Code generation and debugging
- Multimodal understanding
- Efficient local inference
- Competitive performance relative to significantly larger models
Note: Individual benchmark results may vary depending on evaluation methodology, prompt format, inference parameters, quantization, context length, and hardware. Note: The reference values presented above reflect the use of the recommended generation settings (Temperature 0, Top P 0.95, Top K 40).---
⚙️ Recommended Generation Settings
For the best results, change the default sampling configuration to the following values:
| Parameter | Recommended Value |
|---|---|
| Temperature | 0 |
| Top P | 0.95 |
| Top K | 40 |
| Min P | 0.05 |
| Repetition Penalty | 1.1 |
| Presence Penalty | Off |
| Max Tokens | Max |
Recommended Configuration
Temperature: 0.1
Top P: 0.95
Top K: 25
Min P: 0.05
Repetition Penalty: 1.1
Presence Penalty: Off
Max Tokens: Max
⚠️ Important: For optimal output quality, configure your inference interface with the sampling parameters above.
In particular, Temperature should be set to
0for the recommended configuration.Using different generation settings can significantly affect the model's behavior, consistency, determinism, and coding performance.
💻 Recommended Use Cases
Nexora-Gemma-4-E4B-Coder is designed for:
- Code generation
- Code completion
- Code explanation
- Debugging
- Refactoring
- Software architecture
- Technical reasoning
- Technical documentation
- Programming assistance
- Repository analysis
- Error analysis
- Screenshot-based debugging
- Multimodal code understanding
- Visual programming interfaces
👁️ Vision / Multimodal Usage
The model supports multimodal inference through the provided multimodal projector.
For vision-enabled inference, use:
gemma-4-e4b-it.BF16-mmproj.gguf
This enables supported multimodal inference workflows involving:
- Screenshots
- User interfaces
- Programming environments
- Visual debugging
- Diagrams
- Technical images
- Code displayed in images
🦙 Ollama Note for Vision Models
⚠️ Important: Ollama currently does not support separate
mmprojfiles for vision models in the same way asllama.cpp.
To create an Ollama model from a vision-enabled model, place your Modelfile in the same directory as the appropriate merged BF16 model and run:
ollama create nexora-gemma-4-e4b-coder -f ./Modelfile
Replace nexora-gemma-4-e4b-coder with your preferred model name.
🛠️ Built With Unsloth
This model was fine-tuned and converted using Unsloth.
Unsloth provides optimized workflows for efficient model fine-tuning and conversion.
📄 Model Information
| Property | Value |
|---|---|
| Model | Nexora-Gemma-4-E4B-Coder |
| Base Family | Gemma 4 |
| Base Variant | E4B |
| Specialization | Coding |
| Format | GGUF |
| Architecture | Multimodal Language Model |
| Fine-Tuning | Nexora |
| Conversion | Unsloth |
| Inference | llama.cpp compatible |
⚠️ Disclaimer
This model is provided for research, development, experimentation, and local inference.
Actual performance may vary depending on:
- Quantization method
- Inference backend
- Hardware
- Context length
- Prompt formatting
- Sampling parameters
- Evaluation methodology
Generated code should always be reviewed and tested before being used in production environments.
⭐ Nexora
Intelligence. Code. Evolve.
Built for developers who want capable AI coding assistance running locally.