Lux-V2-Pro
Lux-V2-Pro is a Korean-enhanced, fully fine-tuned LLM built on top of google/gemma-4-31B-it by PoSTMEDIA AI Lab, succeeding Lux-V1-Pro.
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. As the maximum-capability tier of the Lux line, V2-Pro improves on Lux-V1-Pro across the board while adding substantial Korean domain knowledge.
Highlights
- Upgrade over Lux-V1-Pro on 16 of 18 internal benchmarks — including AIME 2024 93.3 (+13.3) and AIME 2025 +13.4
- 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 — general knowledge (MMLU 86.6) and instruction following (IFEval 93.7) end above the V1-Pro level
- Dense 31B — the strongest reasoning tier of the Lux family
- Verified training data — synthetic datasets produced with execution- and rule-based verification pipelines
Model Overview
| Specification | Details |
|---|---|
| Base Model | google/gemma-4-31B-it |
| Parameters | 31B (dense) |
| Architecture | Decoder-only Transformer (dense) |
| 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-Pro
All results below were measured in-house under a single unified protocol (identical prompts, sampling, and generation budgets for both models).
| Benchmark | gemma-4-31B-it (base) | Lux-V1-Pro | Lux-V2-Pro |
|---|---|---|---|
| MMLU | 86.6 | 85.9 | 86.6 |
| AIME 2024 | 83.3 | 80.0 | 93.3 |
| AIME 2025 | 66.7 | 63.3 | 76.7 |
| HMMT 2025 | 70.0 | 60.0 | 63.3 |
| IFEval | 93.2 | 92.4 | 93.7 |
| KMMLU | 76.6 | 74.8 | 76.9 |
| KMMLU-Pro | 78.2 | 77.0 | 77.4 |
| CLIcK | 85.1 | 84.8 | 85.5 |
| KoBALT | 73.4 | 71.9 | 73.9 |
| HAE-RAE Bench | 82.0 | 80.6 | 83.1 |
| HRM8K | 85.8 | 84.7 | 85.8 |
| KoSimpleQA | 91.2 | 90.0 | 90.8 |
| KoSQA-EM | 38.0 | 36.1 | 38.5 |
| Average (17 common benchmarks) | 76.6 | 75.5 | 78.2 |
The full 18-benchmark suite shows 16 wins, 1 tie (GPQA), and 2 minor regressions (GSM8K −1.2, IFBench −0.2) versus Lux-V1-Pro. Averaged over the 17 benchmarks measured for all three models, Lux-V2-Pro scores 78.2 — above both the original Gemma-4 base (76.6) and Lux-V1-Pro (75.5). (MMLU-Pro was not measured for the base model due to generation-budget constraints.)
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-Pro"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype=torch.bfloat16,
device_map="auto",
)
messages = [{"role": "user", "content": "문화재보호법의 목적을 두 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
- The most demanding Korean reasoning and generation workloads in the Lux line
- Korean domain QA (law, cultural heritage, tourism) with source-grounded knowledge
- Competition-level mathematical assistance
- Drop-in upgrade for existing Lux-V1-Pro 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{lux2026pro,
title = {Lux-V2-Pro: Capability-Preserving Korean Domain Adaptation of Gemma-4},
author = {{PoSTMEDIA AI Lab}},
year = {2026},
url = {https://huggingface.co/PoSTMEDIA/Lux-V2-Pro}
}
Contact
Questions and feedback — please open a discussion on the model page.