GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
PoSTMEDIA
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
PoSTMEDIA
Model Tree
Qwen/Qwen3.8-27B
Input Modalities
Output Modalities
Supported Functionality
<think> traces| Specification | Details |
|---|---|
| Base Model | Qwen/Qwen3.8-27B |
| Parameters | 27B (hybrid attention: 16 full + 48 linear layers) |
| Training Precision | BF16 |
| Inference Precision | BF16 |
| Context Length | Inherits from Qwen3.8 base |
| Fine-Tuning Method | Full-parameter SFT (Capability-Preserving recipe) |
| Mode | Non-thinking (direct answer) |
| Languages | Korean, English |
All results measured in-house under a single unified protocol (identical prompts, official non-thinking sampling — temperature 0.7, top-p 0.8, top-k 20, presence penalty 1.5 — and identical generation budgets). Competition-math rows (AIME/HMMT) are the mean of 3 sampling seeds for both models to suppress small-sample noise.
| Benchmark | Qwen3.8-27B (base) | Xin-V1 |
|---|---|---|
| MMLU | 84.5 | 83.5 |
| MMLU-Pro | 82.1 | 82.0 |
| GPQA | 80.3 | 81.3 |
| GSM8K | 86.1 | 86.8 |
| AIME 2024† | 77.8 | 83.3 |
| AIME 2025† |
† Competition-math rows are the mean of 3 sampling seeds for both models. ‡ Judge-scored short-answer QA; Xin-V1's more direct answer style contributes to this gain (exact-match on the same set is comparable to base).
Xin-V1 is trained on PoSTMEDIA's in-house Korean synthetic data assets, generated and quality-controlled by our internal data factory:
Correctness of the synthetic data is enforced by verification gates (code execution, symbolic math equivalence, rule checkers) 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.
bash
pip install transformers acceleratepython
from transformers import AutoModelForCausalLM, AutoTokenizerimport torch model_name = "PoSTMEDIA/Xin-V1"tokenizer = AutoTokenizer.from_pretrained(model_name)model = AutoModelForCausalLM.from_pretrained( model_name, torch_dtype=torch.bfloat16, device_map="auto",) messages = [{"role": "user", "content": "한국의 전통 명절 세 가지를 소개해줘."}]inputs = tokenizer.apply_chat_template( messages, add_generation_prompt=True, enable_thinking=False, return_tensors="pt",).to(model.device)outputs = model.generate( inputs, max_new_tokens=512, do_sample=True, temperature=0.7, top_p=0.8, top_k=20,)print(tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True))Note: Xin-V1 is tuned for non-thinking use. Pass
enable_thinking=Falsetoapply_chat_template(as above) and use the official non-thinking sampling parameters (temperature 0.7, top-p 0.8, top-k 20, presence penalty 1.5).
bibtex
@misc{xin2026, title = {Xin-V1: Capability-Preserving Korean Fine-Tuning of Qwen3.8}, author = {{PoSTMEDIA AI Lab}}, year = {2026}, url = {https://huggingface.co/PoSTMEDIA/Xin-V1}}Questions and feedback — please open a discussion on the model page.
pip install transformers accelerate| 71.1 |
| 71.1 |
| AIME 2026† | 78.9 | 81.1 |
| HMMT 2025† | 56.7 | 54.4 |
| IFEval | 86.7 | 83.7 |
| IFBench | 71.9 | 74.5 |
| KMMLU | 72.1 | 70.9 |
| KMMLU-Pro | 65.8 | 66.5 |
| CLIcK | 77.5 | 76.0 |
| KoBALT | 51.1 | 47.1 |
| HAE-RAE Bench | 79.9 | 78.4 |
| HRM8K | 83.5 | 83.3 |
| KoSimpleQA‡ | 56.4 | 66.5 |
| KoSQA-EM | 16.3 | 15.7 |
| Average (all 18) | 71.0 | 71.5 |