ginigen-ai/Rogue-28B-MIX

VISIONConcurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Apr 27, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

ginigen-ai/Rogue-28B-MIX is a 28 billion parameter multimodal language model developed by ginigen-ai, specializing in Korean language reasoning and knowledge. It is a merge of Darwin-28B-KR and QuettaLLMs-27B-Koreasoner-V3, further fine-tuned on K-AI domain data. This model retains multimodal capabilities and the ability to generate reasoning traces, excelling in specific Korean knowledge domains like healthcare and history.

Loading preview...

Rogue-28B-MIX: Korean Reasoning and Multimodal LLM

ginigen-ai/Rogue-28B-MIX is a 28 billion parameter multimodal model specifically designed for Korean language tasks, combining strong reasoning capabilities with extensive Korean knowledge. It is a unique blend, inheriting reasoning strengths from the Darwin-28B-KR lineage and deep Korean K-AI knowledge from the QuettaLLMs-27B-Koreasoner-V3 model, which previously ranked first on the K-AI Leaderboard.

Key Capabilities and Features

  • Hybrid Architecture: Created by merging weights from two specialized Korean models, followed by additional Supervised Fine-Tuning (SFT) on K-AI domain data.
  • Korean Language Expertise: Optimized for Korean reasoning and knowledge, particularly in specialized fields.
  • Multimodal Support: Retains multimodal heads, allowing for potential future multimodal applications.
  • Reasoning Trace: Preserves the <think> reasoning trace mechanism, aiding in understanding the model's thought process.
  • Performance: Achieves competitive results on Korean benchmarks, notably surpassing its 'parent' model (QuettaLLMs-27B-Koreasoner-V3) in key K-AI Leaderboard categories such as KMMLU History (48%) and KMMLU Health (81%).
  • Context Length: Features an 8K token context length, which is extensible.
  • License: Released under the Apache 2.0 License.

Ideal Use Cases

  • Korean-centric AI Applications: Excellent for applications requiring deep understanding and generation in Korean.
  • Specialized Knowledge Domains: Particularly strong in healthcare and history within the Korean context.
  • Reasoning-intensive Tasks: Benefits from its preserved reasoning trace for complex problem-solving.
  • Research and Development: Suitable for researchers exploring hybrid model architectures and domain-specific fine-tuning for Korean LLMs.

Finetunes
1 models