Monibee-Fudgekins/gemma-coder

VISIONConcurrent Unit Cost:2Model Size:26BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 29, 2026License:gemmaArchitecture:Transformer Featherless Exclusive Cold

Monibee-Fudgekins/gemma-coder is a coding-focused QLoRA fine-tune of Google's Gemma 4 26B A4B-it, an MoE model with approximately 4 billion active parameters. Specialized for code generation and assistance, it excels in Python, JavaScript/React, Go, and Java. This model is continuously trained weekly and is intended for use within a sandboxed agent for code development.

Loading preview...

Monibee-Fudgekins/gemma-coder: Coding-Focused Gemma Fine-tune

This model is a QLoRA fine-tune of google/gemma-4-26B-A4B-it, an MoE (Mixture of Experts) architecture with roughly 4 billion active parameters. Developed by Monibee-Fudgekins, it is specifically designed for coding assistance and serves as the default agent model for the remote-agent-dev-platform.

Key Capabilities & Features

  • Specialized for Coding: Optimized for code generation and assistance across multiple languages.
  • Multi-language Support: Proficient in Python, JavaScript/React, Go, and Java.
  • Continuous Training: Undergoes weekly retraining cycles using QLoRA (Unsloth) on Kaggle's free T4 GPUs, ensuring ongoing improvement.
  • Evaluation: Achieves a 100.00% pass@1 rate on its sandboxed multi-language harness for executed problems, demonstrating strong functional correctness.
  • Training Data: Utilizes the bigcode/self-oss-instruct-sc2-exec-filter-50k dataset.

Intended Use Cases

  • Code Generation: Ideal for generating code snippets and functions in supported languages.
  • Coding Assistance: Useful for developers seeking help with programming tasks within a sandboxed environment.
  • Agent Integration: Designed to be integrated into development platforms that can run and test generated code.

Limitations

As a smaller model trained on free-tier resources, it may produce incorrect or insecure code. It is not intended for safety-critical applications or for running generated code without review and testing. Quality is directly tied to the evolving training data.