mkd-ai/Keural-Nova-1.1
Keural Nova 1.1 is an instruction-tuned 35B-parameter Mixture-of-Experts model by MKD, based on Qwen/Qwen3.6-35B-A3B with ~3B active parameters. It is specifically fine-tuned for Korean conversational quality and comprehension, showing substantial improvements on Korean benchmarks like KoBEST. While excelling in Korean dialogue, it maintains some English capabilities and supports extended contexts up to 1,000,000 tokens via YaRN.
Loading preview...
Keural Nova 1.1: Korean-Optimized Instruction Model
Keural Nova 1.1, developed by MKD, is an instruction-tuned model built upon Alibaba Cloud's Qwen3.6-35B-A3B, a 35-billion parameter Mixture-of-Experts (MoE) architecture with approximately 3 billion active parameters per token. This version is the result of a supervised fine-tune (LoRA, merged into base weights) specifically targeting Korean conversational quality and comprehension, while also incorporating an English replay mix to mitigate regression in general capabilities.
Key Capabilities & Performance
- Korean Language Proficiency: Achieves a significant improvement of +4.41 points on the KoBEST benchmark compared to its base model, making it highly suitable for Korean dialogue and text understanding.
- Context Length: Inherits the base model's native support for 262,144 tokens and has been tested in production with up to 1,000,000 tokens using YaRN rope scaling.
- Answer Formatting: Substantially improves answer formatting, particularly noted in benchmarks like GSM8K.
Intended Use & Limitations
- Good for: Korean conversational AI, Korean text comprehension and generation, bilingual (Korean/English) assistant applications, and RAG-style grounded answering in Korean.
- Not suitable for: Code generation, as it shows a significant regression in HumanEval performance (from 62.2 to 11.6 pass@1). There are also minor regressions (1-2 points) on knowledge-heavy benchmarks like KMMLU, HAE-RAE, and MMLU compared to the base model.
Keural Nova 1.1 represents the full training endpoint (2 epochs) and is offered alongside v1.0, which presents slightly different trade-offs in benchmark scores.