FINAL-Bench/Ourbox-35B-JGOS

TEXT GENERATIONConcurrent Unit Cost:3Model Size:35.1BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 10, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Ourbox-35B-JGOS is a 35-billion-parameter Mixture-of-Experts (MoE) reasoning model developed by FINAL-Bench / VIDRAFT_LAB, built upon the Qwen3.6-35B-A3B backbone. This model is specialized for Korean language tasks through a Darwin evolutionary FFN-level merge, while maintaining strong performance in English graduate-level science, achieving 86.36% on GPQA Diamond. It features a hybrid attention architecture, 262K context length, and multi-token prediction, making it suitable for advanced multilingual reasoning and generation.

Loading preview...

Ourbox-35B-JGOS: Korean-Specialized Reasoning MoE

Ourbox-35B-JGOS is a 35-billion-parameter Mixture-of-Experts (MoE) model from FINAL-Bench / VIDRAFT_LAB, created using their Darwin evolutionary breeding platform. It leverages the Qwen3.6-35B-A3B backbone and is uniquely specialized for Korean language through an FFN-level evolutionary merge, rather than traditional retraining.

Key Capabilities & Features

  • Korean Specialization: Enhanced Korean reasoning, comprehension, and generation, including robust handling of scientific/technical text.
  • Advanced Reasoning: Achieves 86.36% on GPQA Diamond (English graduate-level science questions), outperforming its Qwen3.6-35B-A3B backbone (86.0%) and models like GLM-5.1 (86.2%) with only ~3B active parameters.
  • Darwin Evolutionary Merge: Utilizes a novel FFN-level recombination of weight tensors from specialized donor models, preserving the backbone's architecture while adding targeted capabilities.
  • Hybrid Attention & Long Context: Inherits Qwen3.6-35B-A3B's hybrid attention (ΒΎ linear + ΒΌ full) and a 262K token context length, enabling efficient processing of very long inputs.
  • Multilingual Support: While Korean-specialized, it retains strong capabilities in English, Chinese, Japanese, and other languages.
  • Thinking Model: Designed to produce detailed reasoning traces (<think>...</think>) before generating a final answer, beneficial for complex problem-solving.

Ideal Use Cases

  • Applications requiring high-quality Korean language processing and generation.
  • Tasks demanding advanced reasoning capabilities in both Korean and English.
  • Scenarios where long context understanding is critical.
  • Developers seeking a specialized model that improves on its backbone without full retraining.

Finetunes
2 models