mkd-hossain/Keural-Nova-v1.2-experimental

TEXT GENERATIONPricing:Input $0.4 / Cached $0.07 / Output $4Concurrent Unit Cost:3Model Size:35.1BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 30, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Keural Nova v1.2-experimental by MKD is a 35.1 billion parameter bilingual (Korean/English) large language model, fine-tuned from Qwen3.6-35B-A3B. It is specifically optimized for superior tool-calling capabilities, enhanced Korean fluency, and improved code generation, while maintaining general language understanding. This model excels in agentic workflows, function calling, and coding tasks, particularly within a 256K native context window.

Loading preview...

Keural Nova v1.2-experimental: Bilingual LLM for Agents and Code

Keural Nova v1.2-experimental, developed by MKD, is a 35.1 billion parameter bilingual (Korean/English) large language model. It is a LoRA supervised fine-tune of Qwen3.6-35B-A3B, designed to significantly improve Korean fluency, tool-calling, and code generation while preserving the base model's general capabilities. This experimental release represents the strongest Keural Nova version to date in these specialized areas.

Key Capabilities & Performance Highlights

  • Best-in-line Tool-Calling: Achieves 12/12 on MKD's function-calling suite, demonstrating robust multi-turn capabilities and verified tool-calling in very long contexts (~237k tokens). It uses Qwen XML for tool calls.
  • Enhanced Code Generation: Recovers and surpasses the base model with a HumanEval score of 68.3, a significant improvement over previous Nova versions.
  • Superior Korean Fluency: Achieves the highest scores among all Keural Nova versions on KoBEST (70.0) and KMMLU (63.5).
  • Native 256K Context: Leverages the Qwen3.6 architecture for a native 256K token context window, recommended for agentic and tool-calling workloads.

Intended Use Cases

  • General Assistant: Capable of handling diverse conversational tasks.
  • Korean/English Chat and RAG: Optimized for bilingual interactions and retrieval-augmented generation.
  • Agentic / Tool-Calling Workloads: Ideal for web search, document QA, and function calling, especially in long-context scenarios.
  • Coding: Strong performance in code generation and understanding.
  • Korean Enterprise Assistants: Particularly suited for applications requiring high fluency and agent capabilities in Korean contexts.

While showing minor trade-offs in GSM8K math word problems and small knowledge dips on MMLU/HAE-RAE compared to the base, these are expected SFT trade-offs. The model is released under the Apache-2.0 license.

Finetunes
3 models