DireDreadlord/GemCod-Zircon-270M-XM

TEXT GENERATIONConcurrent Unit Cost:1Model Size:0.3BQuant:BF16Context Size:32kPublished:Jun 13, 2026License:gemmaArchitecture:Transformer0.0K Featherless Exclusive Cold

DireDreadlord/GemCod-Zircon-270M-XM is a 270 million parameter Gemma3-based causal language model, fine-tuned for specialized code generation tasks. This model excels at Fill-In-the-Middle (FIM) code completion and multi-turn chat interactions, acting as an embedded agent for code development. Its small size allows efficient operation on laptop-grade GPUs, though its experimental nature and specialized training may lead to less accurate long-form general coding outputs.

Loading preview...

Model Overview

DireDreadlord/GemCod-Zircon-270M-XM is a compact 270 million parameter model built on the Gemma3 architecture, specifically fine-tuned from google/gemma-3-270m-it. Unlike the standard GemCod models, Zircon features custom templating for Fill-In-the-Middle (FIM) tasks and multi-turn chat capabilities, positioning it as an experimental embedded agent for code assistance.

Key Capabilities

  • FIM Code Completion: Designed to complete missing code segments within existing structures.
  • Multi-turn Chat: Supports interactive conversations for code generation and debugging.
  • Lightweight: Its small size enables comfortable execution on laptop-grade GPUs.

Training Details

The model was trained for 1,500 steps using Supervised Fine-Tuning (SFT) on the Etherll/code-fim-v2 dataset, which comprises approximately 64,000 rows templated with a custom .jinja chat format. This specialized training focuses on agentic architecture within the GemCod family.

Limitations

  • Experimental Variant (XM): Trained for only 1,500 steps, which may result in slightly incorrect outputs for long-form generation.
  • Specialized Focus: Its specialized formatting and dataset mean it has inferior generation capabilities for general coding tasks compared to models optimized for broader code generation.