mkd-ai/Keural-Nova-v1.1

TEXT GENERATIONConcurrent Unit Cost:3Model Size:35.1BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 16, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Keural Nova v1.1 by mkd-ai is a 35B-parameter Mixture-of-Experts (MoE) model, based on Qwen/Qwen3.6-35B-A3B, with approximately 3B active parameters per token. This instruction-tuned model is specifically optimized for high-quality Korean conversational AI and text comprehension, offering significant improvements in Korean language tasks. It supports a context length of up to 8,192 tokens and is intended for bilingual (Korean/English) assistant use and RAG-style grounded answering in Korean.

Loading preview...

Keural Nova v1.1: Korean-Focused Instruction-Tuned MoE Model

Keural Nova v1.1, developed by mkd-ai, is an instruction-tuned model built upon the Qwen/Qwen3.6-35B-A3B base architecture. This 35B-parameter Mixture-of-Experts (MoE) model, with ~3B active parameters per token, has been fine-tuned to excel in Korean conversational quality and text comprehension, while maintaining some English capabilities.

Key Capabilities & Performance

  • Korean Language Proficiency: Substantially improves Korean conversational AI and text comprehension, as evidenced by a +4.41 score increase on the KoBEST benchmark compared to its base model.
  • Bilingual Support: Designed for effective bilingual (Korean/English) assistant use.
  • Context Window: Supports sequences up to 8,192 tokens, inheriting the base model's 262,144-token architectural context window (though long context behavior is untested).
  • Robust Formatting: Demonstrates significant improvements in answer formatting, particularly for tasks like GSM8K.

Intended Use Cases

  • Korean Conversational AI: Ideal for chatbots and virtual assistants requiring high-quality Korean dialogue.
  • Korean Text Processing: Excellent for comprehension and generation of Korean text.
  • RAG Systems: Suitable for RAG-style grounded answering in Korean.

Important Considerations & Limitations

  • Not for Code Generation: This model is explicitly not suitable for code generation, showing a significant regression in HumanEval performance due to corrupted formatting in training data. Users needing code generation should use the base Qwen model.
  • Knowledge-Heavy Tasks: Exhibits slight regressions (1-2 points) on general knowledge-recall benchmarks (KMMLU, HAE-RAE, MMLU) compared to the base model.
  • Text-Only Fine-tune: While the base architecture is multimodal, this fine-tune focused solely on text; vision capabilities were neither trained nor evaluated.

Finetunes
1 models