janPaje/iolai-gemma4-hybrid

TEXT GENERATIONConcurrent Unit Cost:1Model Size:12BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 12, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The janPaje/iolai-gemma4-hybrid is a 12 billion parameter language model based on the Gemma4 architecture, designed for numeral solving tasks. It employs a unique three-pass pipeline, combining a symbolic numeral solver with a budget-managed LLM to accurately process and answer problems involving numeral systems. This model excels at tasks requiring precise numerical reasoning and translation, particularly when a consistent numeral system can be identified or when falling back to an LLM for complex cases.

Loading preview...

janPaje/iolai-gemma4-hybrid: A Hybrid Numeral Solver

The janPaje/iolai-gemma4-hybrid model, developed by janPaje, is a 12 billion parameter language model built upon the Gemma4 architecture. It distinguishes itself through a novel three-pass pipeline designed to efficiently and accurately solve problems involving numeral systems.

Key Capabilities

  • Symbolic Numeral Solving: The initial pass utilizes a numeral_solver.py component to brute-force and identify consistent numeral systems (base, word values, word-order convention) from problem examples, providing exact text_to_num and num_to_text conversions.
  • Hybrid Approach: Problems not solvable symbolically are passed to the Gemma4-12B LLM for a baseline answer.
  • Budget-Managed Reasoning: An "Upgrade pass" refines baseline answers using time-sliced reasoning and task-specific method prompts, replacing answers only when the upgrade parses correctly. This approach prioritizes efficiency, with truncated reasoning on Gemma4-12B proving more effective for short answers.
  • Resource Efficiency: The model loads 4-bit quantized weights via bitsandbytes at startup, enabling it to run on hardware with limited VRAM (e.g., 16 GB T4 GPUs).

Good for

  • Tasks requiring precise numeral system identification and conversion.
  • Applications where a hybrid symbolic-LLM approach can enhance accuracy and efficiency for numerical problems.
  • Scenarios demanding budget-managed LLM inference, optimizing for speed and cost while maintaining answer quality.

Finetunes
1 models