ludx/gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2

TEXT GENERATIONPricing:Input $1.2 / Cached $0.24 / Output $4.8Concurrent Unit Cost:1Model Size:12BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 10, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

ludx/gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2 is a 12 billion parameter Gemma 4 model fine-tuned by ludx for advanced coding and agentic workflows. This model excels at multi-step technical tasks, tool use, and debugging, demonstrating a 3.5x improvement over its base model on agentic benchmarks. It is designed for developers needing a robust local coding and agentic worker, supporting a 32768 token context length.

Loading preview...

Model Overview

ludx/gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2 is a full-precision (bf16) Gemma 4 12B model, specifically fine-tuned for coding and agentic capabilities. This v2 release significantly enhances the model's ability to read, reason, use tools, and execute multi-step technical tasks. It is provided as safetensors master weights, intended for builders to create custom quants, further fine-tune, or run directly in transformers.

Key Capabilities & Differentiators

  • Agentic Performance: Achieves approximately 3.5x higher scores on the tau2-bench telecom benchmark (around 55% vs. 15% for the base model), indicating strong performance in diagnose-fix-verify loops typical of terminal and debugging work.
  • Coding & Tool Use: Designed for writing code, running commands, and structured tool-calls using Gemma 4's native protocol. It grounds its actions, avoiding fabrication in coding/terminal tasks.
  • Multi-step Reasoning: Incorporates real multi-step tool-use trajectories (read → reason → act → verify) and verified chain-of-thought over Python tasks.
  • Specialized Focus: Trades some general knowledge breadth (scoring slightly below the base on MMLU-Pro) for deep expertise in coding and agentic problem-solving.
  • Context Length: Supports a context window of 32768 tokens.

Use Cases

  • Local Coding Assistant: Ideal for developers requiring a powerful local model for code generation, debugging, and technical problem-solving.
  • Agentic Workflows: Suitable for applications involving tool use, multi-step task execution, and automated technical operations.
  • Further Fine-tuning: Serves as a clean base for continued training or LoRA adaptations due to its full-precision weights.

This model is English-centric and specialized for its target domains, with reduced refusals due to its task-focused training.

Finetunes
1 models