PoSTMEDIA/Xin-V1
PoSTMEDIA/Xin-V1 is a 27-billion parameter, Korean-enhanced, instruction-tuned large language model built on Qwen/Qwen3.8-27B. Developed by PoSTMEDIA AI Lab using a Capability-Preserving Full Fine-Tuning methodology, it excels in direct answer generation without reasoning traces, with significantly improved competition math performance and strong Korean-first tuning. This model is optimized for fast, direct responses in Korean and English, particularly for mathematical, coding, and general instruction-following tasks.
Loading preview...
Xin-V1: Korean-Enhanced, Direct-Answer LLM
Xin-V1 is a 27-billion parameter, instruction-tuned large language model developed by PoSTMEDIA AI Lab, based on the Qwen/Qwen3.8-27B architecture. It is the first model in the Xin line, distinguished by its Korean-first tuning and a unique Capability-Preserving Full Fine-Tuning methodology. This approach ensures deep domain adaptation without eroding the base model's reasoning, instruction-following, and multilingual abilities.
Key Capabilities
- Korean-First Tuning: Trained extensively on PoSTMEDIA's in-house verified Korean synthetic datasets.
- Non-Thinking Product: Designed for fast, direct answers without emitting reasoning traces, making it efficient for immediate responses.
- Enhanced Competition Math: Shows significant improvement in AIME and HMMT benchmarks over its base model.
- Base Capability Preservation: Maintains strong performance in general knowledge (MMLU, GPQA) and instruction following.
- Hybrid-Attention Architecture: Utilizes a 27B hybrid-attention design (16 full-attention + 48 linear-attention layers) for efficient long-context serving.
Good For
- Korean-first assistants requiring rapid, direct answers.
- Mathematical and coding assistance in both Korean and English.
- General Korean instruction-following workloads where direct responses are preferred.