naazimsnh02/FabGemma

TEXT GENERATIONConcurrent Unit Cost:1Model Size:12BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 16, 2026License:gemmaArchitecture:Transformer0.0K Featherless Exclusive Cold

FabGemma-12B by naazimsnh02 is a 12 billion parameter instruction-tuned Gemma 4 model, specifically optimized for advanced agentic coding, autonomous task planning, and rigorous debugging workflows. It leverages a massive 256K token context window and was fine-tuned on 15.2 million tokens from high-tier coding agent sessions. This model excels at reasoning and planning before acting, making it highly efficient for complex code-related tasks.

Loading preview...

Model Overview

FabGemma-12B is an advanced, reasoning-first optimization of Google's Gemma 4 12B Instruct model, developed by naazimsnh02. It has been specifically fine-tuned to integrate advanced agentic coding, autonomous task planning, and rigorous debugging workflows into its instruction-following capabilities. The model's core habit is to reason and plan before acting, achieved through supervised fine-tuning (SFT) on complex agentic traces.

Key Capabilities & Features

  • Reasoning-First Approach: Modeled after complex, multi-step debugging and tool-use reasoning paths.
  • Base Architecture: Built upon google/gemma-4-12B-it (Dense Transformer).
  • Massive Context Window: Inherits Gemma 4's native 256K token context window.
  • Efficiency: Trained using LoRA, modifying only 2.15% (~262M parameters) of the total network.
  • Specialized Training: Fine-tuned on 15.2 million tokens from high-tier coding agent sessions, primarily from Glint-Research/Fable-5-traces.
  • Structured Output: Organizes outputs into clear, cognitive steps using explicit XML-style formatting for thought processes and tool calls.

Performance & Evaluation

FabGemma-12B demonstrates significant improvements over its base model in agentic workflows. On 105,525 unseen response tokens from 100 held-out coding traces, it achieved:

  • Evaluation Loss: 0.737 (a 53.4% improvement over the base model's 1.580).
  • Perplexity: 2.089 (a 57.0% improvement over the base model's 4.856).

Ideal Use Cases

  • Agentic Coding: Tasks requiring advanced code generation, planning, and execution.
  • Autonomous Task Planning: Developing multi-step plans for complex problems.
  • Debugging Workflows: Assisting in identifying and proposing fixes for code issues.

Limitations

  • Specialized Focus: Optimized for code architecture, script execution planning, and debugging; general trivia or encyclopedic knowledge may not be its strength.
  • Modality: Strictly a text-to-text model, lacking core vision or audio capabilities.
  • Language: Primarily fine-tuned on English-centric environments.
  • Inherited Biases: Inherits safety baselines, core biases, and underlying assumptions from the original google/gemma-4-12B-it foundation.

Finetunes
2 models