mkd-ai/Keural-Nova-1.0

TEXT GENERATIONPricing:Input $0.4 / Cached $0.07 / Output $4Concurrent 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 1.0 by mkd-ai is a 35B-parameter Mixture-of-Experts (MoE) instruction-tuned model, with approximately 3B active parameters per token, built upon Qwen/Qwen3.6-35B-A3B. This model is specifically fine-tuned for high-quality Korean conversational AI and text comprehension, significantly improving performance on Korean benchmarks like KoBEST. It supports extended context lengths up to 1,000,000 tokens via YaRN, making it suitable for long-document processing in Korean.

Loading preview...

Keural Nova 1.0: Korean-Focused Instruction-Tuned MoE Model

Keural Nova 1.0, developed by mkd-ai, is an instruction-tuned model based on the Qwen/Qwen3.6-35B-A3B architecture, a 35-billion parameter Mixture-of-Experts (MoE) model. It features approximately 3 billion active parameters per token. This version is the result of a supervised fine-tune (LoRA) specifically targeting Korean conversational quality and comprehension, while including an English replay mix to mitigate capability regression.

Key Capabilities & Differentiators

  • Superior Korean Performance: Achieves a significant improvement on the KoBEST benchmark (+5.94 points over the base model), making it highly effective for Korean dialogue and text understanding.
  • Extended Context Window: Inherits the base model's capability for up to 262,144 tokens natively and supports up to 1,000,000 tokens via YaRN rope scaling, validated in production for long-context applications.
  • Optimized for Conversational AI: Fine-tuned with a focus on Korean conversational AI, text comprehension, and bilingual (Korean/English) assistant use.
  • Robust Training: Trained on a diverse dataset of 253,528 examples, predominantly Korean, using a LoRA method that adapted attention, Gated-DeltaNet, and all 256 routed experts per layer.

When to Use Keural Nova 1.0

  • Korean Conversational AI: Ideal for chatbots, virtual assistants, and dialogue systems requiring high-quality Korean interaction.
  • Korean Text Comprehension & Generation: Excels in tasks involving understanding and generating Korean text.
  • Bilingual (Ko/En) Assistant Use: Suitable for applications requiring seamless switching between Korean and English.
  • RAG-style Grounded Answering in Korean: Effective for retrieval-augmented generation scenarios where answers need to be grounded in Korean documents.

Important Considerations

  • Not for Code Generation: This model shows a significant regression in code generation performance (HumanEval pass@1 drops from 62.2 to 20.7) due to training data issues. For coding tasks, the base Qwen model is recommended.
  • Minor Knowledge Regression: Exhibits small regressions (1-2 points) on knowledge-heavy benchmarks like KMMLU, HAE-RAE, and MMLU compared to the base model.
  • Vision Untouched: While inheriting a multimodal architecture, the vision tower was neither trained nor evaluated; treat it as a text-only model.

Finetunes
1 models