--- tags: - gguf - llama.cpp - unsloth - vision-language-model license: mit base_model: - google/gemma-4-E4B-it datasets: - guell00/qwen-3.8-code --- ![image](https://cdn-uploads.huggingface.co/production/uploads/644afe169279988e0cbcd2d9/oNzEYG6aTWqcmX2GMcX3I.png) # 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 --- ## πŸš€ Quick Start ### Text-only inference Run the model with `llama.cpp`: ```bash llama-cli -hf guell00/Nexora-Gemma-4-E4B-Coder --jinja ```` ### Multimodal inference For vision-language inference: ```bash llama-mtmd-cli -hf guell00/Nexora-Gemma-4-E4B-Coder --jinja ``` --- ## πŸ“¦ 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. --- # πŸ“Š 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%** | β€” | **53.0%** | **83.4%** | | **Gemma 4 E4B** | **69.4%** | **42.5%** | **52.0%** | **940** | **58.6%** | **42.2%** | β€” | β€” | **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).--- # βš™οΈ 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 ```text 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 `0`** for the recommended configuration. > > Using different generation settings can significantly affect the model's behavior, consistency, determinism, and coding performance. --- # πŸ’» 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 --- # πŸ‘οΈ Vision / Multimodal Usage The model supports multimodal inference through the provided multimodal projector. For vision-enabled inference, use: ```text 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 --- # πŸ¦™ Ollama Note for Vision Models > ⚠️ **Important:** Ollama currently does not support separate `mmproj` files for vision models in the same way as `llama.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: ```bash ollama create nexora-gemma-4-e4b-coder -f ./Modelfile ``` Replace `nexora-gemma-4-e4b-coder` with your preferred model name. --- # πŸ› οΈ Built With Unsloth This model was fine-tuned and converted using [Unsloth](https://github.com/unslothai/unsloth). Unsloth provides optimized workflows for efficient model fine-tuning and conversion. [](https://github.com/unslothai/unsloth) --- # πŸ“„ 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.