cyzero-kim/gemma-4-31B-Opus-4.6-Reasoning

VISIONPricing:Input $0.48 / Cached $0.1 / Output $1.44Concurrent Unit Cost:2Model Size:31BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Apr 28, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

The cyzero-kim/gemma-4-31B-Opus-4.6-Reasoning model is a 31 billion parameter Gemma-4 based language model, fine-tuned by cyzero-kim with a 32768 token context length. It is specifically optimized for complex, step-by-step logical reasoning tasks, leveraging the Opus-4.6-Reasoning dataset. This model excels at generating structured, analytical thought processes before delivering a final answer, making it suitable for detailed problem-solving and planning.

Loading preview...

Overview

This model, cyzero-kim/gemma-4-31B-Opus-4.6-Reasoning, is a fine-tuned and merged version of the Gemma-4 31B foundation model. It was developed by cyzero-kim, building upon the cloudbjorn/gemma-4-31B-Opus-4.6-Reasoning base, with a primary focus on enhancing reasoning capabilities. The model integrates the Gemma-4's native <|channel> architecture to enforce strict, logical step-by-step reasoning.

Key Capabilities

  • Structured Reasoning: Outputs internal thought processes within <|channel>thought bounds before providing a final answer, enabling transparent and verifiable reasoning.
  • Complex Problem Solving: Designed to act as a deeply analytical agent, capable of planning intricate scenarios like cloud deployments and performing logical deductions.
  • Gemma 4 Architecture Adherence: Strictly follows the Gemma 4 multimodal and reasoning formats.

Training Details

The model was fine-tuned using the Crownelius/Opus-4.6-Reasoning-3300x dataset. While standard knowledge benchmarks like ARC Challenge showed a minor regression (from 69.88% to 69.54% acc_norm) compared to the base Gemma-4-31B, the training significantly improved structural output for reasoning tasks. Training was conducted using the Eschaton Engine (Cloudbjorn) with bfloat16 precision, employing an 8-bit Paged AdamW optimizer over 1 epoch with a 2048 sequence length.

Ideal Use Cases

This model is particularly well-suited for applications requiring:

  • Detailed, multi-step problem-solving.
  • Generation of logical plans or strategies.
  • Tasks where the process of reasoning is as important as the final answer.
  • Scenarios demanding analytical depth, such as infrastructure planning or complex decision support.

Finetunes
1 models