PoSTMEDIA/Lux-V2

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:2Model Size:26BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 11, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

PoSTMEDIA AI Lab's Lux-V2 is a 26 billion parameter Mixture-of-Experts (MoE) large language model, built upon Google's gemma-4-26B-A4B-it. It is specifically fine-tuned for enhanced Korean language capabilities across seven domains, including law, cultural heritage, and mathematics, while preserving and improving the base model's general reasoning and instruction-following abilities. This model is optimized for Korean-first applications requiring strong general reasoning and domain-specific knowledge.

Loading preview...

Lux-V2: Korean-Enhanced MoE LLM

Lux-V2, developed by PoSTMEDIA AI Lab, is a 26 billion parameter Mixture-of-Experts (MoE) model based on google/gemma-4-26B-A4B-it. It 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 skills.

Key Capabilities & Enhancements

  • Significant Upgrade over Lux-V1: Achieves substantial improvements across 16 of 18 internal benchmarks, including notable gains in competition math scores.
  • Integrated Korean Domain Knowledge: Trained on PoSTMEDIA's proprietary Korean synthetic datasets covering general conversation, coding, instruction following, law, cultural heritage, tourism, and mathematics.
  • Preserved and Improved Base Abilities: Unlike typical full fine-tuning, Lux-V2 maintains and often enhances the base model's general capabilities, resulting in a net gain over the unmodified Gemma-4 base model's average benchmark score.
  • MoE Efficiency: Operates with approximately 4 billion active parameters during inference, offering efficiency for its 26 billion total parameters.
  • Verified Training Data: Utilizes synthetic datasets generated with execution- and rule-based verification pipelines to ensure high quality and correctness.

Ideal Use Cases

  • Developing Korean-first AI assistants that require robust general reasoning.
  • Applications demanding source-grounded Korean domain-specific QA in areas like law, cultural heritage, and tourism.
  • Providing mathematical and coding assistance in both Korean and English.
  • As a direct upgrade for existing Lux-V1 deployments seeking enhanced performance and broader Korean domain coverage.

Finetunes
2 models