mkd-hossain/Keural-Nova-v1.1

TEXT GENERATIONPricing:Input $0.4 / Cached $0.07 / Output $4Concurrent Unit Cost:3Model Size:35.1BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 24, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Keural Nova v1.1 is a 35.1 billion parameter instruction-tuned model developed by mkd-hossain, based on Qwen/Qwen3.6-35B-A3B, with 32768 tokens context length. This model is specifically fine-tuned for Korean conversational quality and comprehension, offering substantial improvements in Korean language tasks. It is optimized for applications requiring Korean text generation, understanding, and bilingual (Korean/English) assistant capabilities.

Loading preview...

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

Keural Nova v1.1 is a 35.1 billion parameter instruction-tuned model from mkd-hossain, built upon the Qwen/Qwen3.6-35B-A3B architecture. This version represents the full training endpoint (2 epochs) and is primarily focused on enhancing Korean conversational quality and comprehension. It achieves significant improvements in Korean benchmarks like KoBEST, demonstrating a +4.41 point increase over its base model.

Key Capabilities & Features

  • Korean Language Optimization: Substantially improves performance in Korean conversational AI and text comprehension.
  • Bilingual Support: Designed for bilingual (Korean/English) assistant use, with an English replay mix during fine-tuning to limit regression.
  • Extended Context Window: Inherits the base model's ability to support up to 1,000,000 tokens via YaRN rope scaling, with production-proven serving up to 1M tokens.
  • LoRA Fine-tuning: Utilizes LoRA (rank 32) merged into base weights, adapting attention modules and all 256 routed experts per layer.
  • Transparent Versioning: A sibling release, Keural Nova v1.0, is also available, offering users a choice based on specific workload trade-offs.

Intended Use Cases & Limitations

Good for:

  • Korean conversational AI and chatbots.
  • Korean text comprehension and generation.
  • Bilingual (Korean/English) assistant applications.
  • RAG-style grounded answering in Korean.

Not suitable for:

  • Code Generation: Exhibits a significant regression in code generation performance (HumanEval drops from 62.2 to 11.59 pass@1).
  • Knowledge-Heavy QA: Shows small regressions (1-2 points) on knowledge-recall benchmarks like KMMLU, HAE-RAE, and MMLU compared to the base model.
  • Vision Tasks: While inheriting a multimodal architecture, this fine-tune did not train or evaluate vision, so it should be treated as a text-only model.

Finetunes
1 models