tpls/gemma-4-12B-coder-fable5-composer2.5-v1-sft-v5

TEXT GENERATIONConcurrent Unit Cost:1Model Size:12BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 25, 2026License:gemmaArchitecture:Transformer Featherless Exclusive Cold

The tpls/gemma-4-12B-coder-fable5-composer2.5-v1-sft-v5 is a 12 billion parameter Gemma-4 based model, fine-tuned using QLoRA for enhanced code generation and agentic tool use. It demonstrates a 100% gate pass rate for tool-calling in its production path, making it highly reliable for structured tool interactions. This model is designed for local serving via llama.cpp/Ollama or as a base for further fine-tuning, merging, or quantization, excelling in scenarios requiring robust function calling capabilities.

Loading preview...

Model Overview

The tpls/gemma-4-12B-coder-fable5-composer2.5-v1-sft-v5 is a 12 billion parameter model based on the Gemma-4 architecture, specifically fine-tuned for code generation and agentic tool use. It utilizes sft-qlora techniques to improve native tool-calling capabilities.

Key Capabilities

  • Exceptional Tool-Calling: Achieves a 100% gate pass rate for structured tool calls in its production (shim) parse path, making it highly effective for agentic workflows.
  • Code Generation: Built for generating code, suitable for various programming tasks.
  • Fine-tuning Base: Provided as safetensors weights, it can be used as a base for further fine-tuning, merging, or quantization.
  • Local Deployment: Optimized for local serving via llama.cpp or Ollama, with ready-to-serve GGUF quantizations available.

Training Details

The model was trained using QLoRA (4-bit NF4 base, bf16 compute) with an objective to train on assistant turns only. It incorporated a diverse mix of public datasets focused on tool-dense multiturn interactions, reasoning-heavy tasks, and various function-calling scenarios, including Agent-Ark/Toucan-1.5M, Nanbeige/ToolMind, NousResearch/hermes-function-calling-v1, and Salesforce/xlam-function-calling-60k.

Intended Use

This model is ideal for applications requiring reliable tool use and code generation, particularly when deployed locally. Users should validate tool arguments before execution and maintain human oversight for critical applications due to potential for incorrect or fabricated outputs.

Finetunes
6 models