PoSTMEDIA/Lux-V2-Pro

VISIONPricing:Input $0.48 / Cached $0.1 / Output $1.44Concurrent Unit Cost:2Model Size:31BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 11, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

PoSTMEDIA/Lux-V2-Pro is a 31 billion parameter, Korean-enhanced, fully fine-tuned large language model developed by PoSTMEDIA AI Lab, built upon Google's Gemma-4-31B-it. It utilizes a second-generation Capability-Preserving Full Fine-Tuning methodology to integrate deep Korean domain knowledge across seven categories (conversation, coding, instruction, law, culture, tourism, math) while maintaining strong base model reasoning and instruction-following abilities. This model excels in demanding Korean reasoning and generation tasks, offering significant improvements over its predecessor, Lux-V1-Pro, across various benchmarks including MMLU and AIME. Its primary strength lies in providing advanced Korean domain-specific QA and mathematical assistance.

Loading preview...

Lux-V2-Pro: Korean-Enhanced Gemma-4 for Advanced Reasoning

Lux-V2-Pro is a 31-billion parameter, dense decoder-only Transformer model developed by PoSTMEDIA AI Lab. It is a fully fine-tuned version of google/gemma-4-31B-it, specifically enhanced for the Korean language and domain knowledge. This model represents the second generation of PoSTMEDIA's Capability-Preserving Full Fine-Tuning research, a methodology designed to integrate deep domain adaptation without compromising the base model's core reasoning, instruction-following, and multilingual capabilities.

Key Capabilities and Enhancements

  • Superior Performance: Outperforms Lux-V1-Pro on 16 out of 18 internal benchmarks, with notable gains in AIME 2024 (+13.3) and AIME 2025 (+13.4). It also surpasses the original Gemma-4 base model in average performance across 17 common benchmarks.
  • Extensive Korean Domain Knowledge: Trained on PoSTMEDIA's proprietary Korean synthetic datasets covering seven critical domains: general conversation, coding, instruction following, law, cultural heritage, tourism, and mathematics. This data is rigorously verified using execution- and rule-based pipelines.
  • Preserved Base Model Strengths: Maintains and even improves upon the base model's general knowledge (MMLU 86.6) and instruction-following (IFEval 93.7) abilities, demonstrating the effectiveness of its unique fine-tuning approach.
  • Dense 31B Architecture: Provides the strongest reasoning capabilities within the Lux model family.

Ideal Use Cases

  • Handling complex Korean reasoning and generation workloads.
  • Performing source-grounded Korean domain-specific QA in areas like law, cultural heritage, and tourism.
  • Providing advanced mathematical assistance, including competition-level problems.
  • Serving as a direct upgrade for existing Lux-V1-Pro deployments requiring enhanced performance and Korean domain expertise.

Finetunes
1 models