Lux-V2
Lux-V2 is a Korean-enhanced, fully fine-tuned LLM built on top of google/gemma-4-26B-A4B-it by PoSTMEDIA AI Lab, succeeding Lux-V1.
It is produced with the second generation of PoSTMEDIA's in-house Capability-Preserving Full Fine-Tuning research — a training methodology designed so that deep domain adaptation does not erode the reasoning, instruction-following, and multilingual abilities of the base model. Compared to Lux-V1, the V2 generation delivers a large, across-the-board benchmark improvement while adding substantial Korean domain knowledge.
Highlights
- Large upgrade over Lux-V1 — wins on 16 of 18 internal benchmarks, with competition-math gains of +10 to +20 points
- Korean domain knowledge built in — trained on PoSTMEDIA's in-house Korean synthetic datasets spanning seven domains: general conversation, coding, instruction following, law, cultural heritage, tourism, and mathematics, plus the PoSTMEDIA identity dataset
- Base capability preserved and improved — unlike naive full fine-tuning, the V2 methodology keeps (and on most axes improves) the base model's abilities
- MoE efficiency — 26B total / ~4B active parameters at serving time
- Verified training data — the synthetic datasets are produced with execution- and rule-based verification pipelines rather than unfiltered generation
Model Overview
| Specification | Details |
|---|---|
| Base Model | google/gemma-4-26B-A4B-it |
| Parameters | 26B total / ~4B active |
| Architecture | Decoder-only Transformer (MoE) |
| Training Precision | BF16 |
| Inference Precision | BF16 |
| Context Length | Inherits from Gemma-4 base |
| Fine-Tuning Method | Full-parameter SFT (Capability-Preserving recipe, 2nd gen) |
| Languages | Korean, English |
What's New vs Lux-V1
Lux-V1 demonstrated that a full fine-tune could adapt Gemma-4 without destroying it. Lux-V2 goes further: it adapts the model to Korean domains while measurably improving general capability. All results below were measured in-house under a single unified protocol (identical prompts, sampling, and generation budgets for both models).
| Benchmark | gemma-4-26B-A4B-it (base) | Lux-V1 | Lux-V2 |
|---|---|---|---|
| MMLU | 84.2 | 82.0 | 83.8 |
| AIME 2024 | 90.0 | 73.3 | 90.0 |
| AIME 2025 | 73.3 | 60.0 | 80.0 |
| HMMT 2025 | 66.7 | 36.7 | 56.7 |
| IFEval | 89.6 | 88.5 | 91.7 |
| KMMLU | 72.8 | 70.7 | 72.9 |
| KMMLU-Pro | 72.0 | 67.8 | 70.6 |
| CLIcK | 81.3 | 79.8 | 82.0 |
| KoBALT | 68.3 | 59.1 | 65.9 |
| HAE-RAE Bench | 79.7 | 76.8 | 80.1 |
| HRM8K | 83.5 | 79.9 | 83.6 |
| KoSimpleQA | 86.7 | 83.7 | 87.9 |
| KoSQA-EM | 33.6 | 28.7 | 34.1 |
| Average (all 18 benchmarks) | 74.5 | 69.1 | 74.8 |
The full 18-benchmark suite shows 16 wins and 2 minor regressions (GSM8K −1.9, GPQA −1.0) versus Lux-V1. Averaged over all 18 benchmarks, Lux-V2 scores 74.8 — above both the original Gemma-4 base (74.5) and Lux-V1 (69.1) — i.e., the Korean domain adaptation comes with a net gain over the unmodified base model, not a trade-off.
Training Data
The V2 generation is trained on PoSTMEDIA's in-house Korean synthetic data assets, generated and quality-controlled by our internal data factory:
- General conversation — natural Korean multi-topic dialogue
- Coding — execution-verified code generation and explanation
- Instruction following — rule-verifiable Korean constraint-following tasks
- Law — source-grounded Korean legal knowledge QA
- Cultural heritage — source-grounded Korean heritage knowledge QA
- Tourism — source-grounded Korean tourism knowledge QA
- Mathematics — symbolically verified Korean math reasoning
- PoSTMEDIA identity — hand-curated identity dataset
Correctness of the synthetic data is enforced by verification gates (code execution, symbolic math equivalence, rule checkers, and source-grounding checks) rather than by model self-judgment.
The exact training procedure — schedule, module selection, and the post-training consolidation step that preserves base capability — is an internal research method and is not disclosed in detail.
Quick Start
pip install transformers accelerate
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
model_name = "PoSTMEDIA/Lux-V2"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype=torch.bfloat16,
device_map="auto",
)
messages = [{"role": "user", "content": "경복궁의 역사적 의미를 세 문장으로 설명해줘."}]
inputs = tokenizer.apply_chat_template(
messages, add_generation_prompt=True, return_tensors="pt"
).to(model.device)
outputs = model.generate(inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True))
Use Cases
- Korean-first assistants that must retain strong general reasoning
- Korean domain QA (law, cultural heritage, tourism) with source-grounded knowledge
- Mathematical and coding assistance in Korean and English
- Drop-in upgrade for existing Lux-V1 deployments
Safety & Limitations
- The model can generate incorrect or outdated information; verify high-stakes outputs.
- Korean domain knowledge reflects the training data snapshot and may not cover recent changes (e.g., amended laws).
- Inherits the general limitations and usage considerations of the Gemma-4 base model.
Citation
@misc{lux2026,
title = {Lux-V2: Capability-Preserving Korean Domain Adaptation of Gemma-4},
author = {{PoSTMEDIA AI Lab}},
year = {2026},
url = {https://huggingface.co/PoSTMEDIA/Lux-V2}
}
Contact
Questions and feedback — please open a discussion on the model page.