mkd-ai/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) is a 35.1 billion parameter bilingual (Korean/English) large language model developed by MKD. It is a LoRA fine-tune of Qwen3.6-35B-A3B, specifically optimized for tool-calling, code generation, and Korean language fluency. This version features best-in-line tool-calling capabilities, including robust multi-turn and long-context function calls up to 237k tokens, alongside strong performance in coding and Korean benchmarks. It is primarily intended for agentic workloads, Korean/English chat, RAG, and coding applications.

Loading preview...

Keural Nova v1.2 (Experimental) Overview

Keural Nova v1.2 is an experimental bilingual (Korean/English) large language model developed by MKD, an AI company based in South Korea. This 35.1 billion parameter model is a LoRA supervised fine-tune of Qwen3.6-35B-A3B, specifically engineered to enhance Korean fluency, tool-calling capabilities (including long context), and code generation, while maintaining the base model's general knowledge.

Key Capabilities and Performance

  • Tool-Calling Excellence: Achieves 12/12 on MKD's function-calling suite, demonstrating robust multi-turn interactions and verified tool-calling in very long contexts (up to ~237k tokens). It utilizes Qwen XML for tool calls.
  • Improved Code Generation: Scores 68.3 on HumanEval, significantly surpassing the base model's 62.2 and previous Keural Nova versions, due to a fix in code data processing.
  • Enhanced Korean Language: Achieves the highest scores within the Keural Nova line on Korean benchmarks, with KoBEST at 70.0 and KMMLU at 63.5.
  • Native Long Context: Supports a native context length of 256K tokens, leveraging the Qwen3.6 architecture.
  • Balanced Performance: While showing minor knowledge dips in MMLU and HAE-RAE compared to the base model (an expected SFT trade-off), it offers a strong overall balance, though GSM8K math performance is lower than v1.0/v1.1.

Ideal Use Cases

Keural Nova v1.2 is particularly well-suited for:

  • Agentic and Tool-Calling Workloads: Excels in scenarios requiring function calling, web search, and document QA.
  • Bilingual Chat and RAG: Optimized for both Korean and English conversational AI and retrieval-augmented generation.
  • Code Generation: Strong performance makes it suitable for various coding tasks.
  • Korean Enterprise Assistants: Especially valuable for applications in Korean business environments requiring advanced AI capabilities.

Finetunes
3 models