GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
X-LawJapan
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
X-LawJapan
Model Tree
Qwen/Qwen3.8-27B
Input Modalities
Output Modalities
Supported Functionality
All numbers from the paper; same prompts and retrieval for every model; baselines one run each. To reproduce or compare your own model, use XLaw-Bench.
| Benchmark | This model | Gemma 4 12B | LLM-jp-4.1 33B (thinking) | Qwen3-Swallow 32B (thinking) |
|---|---|---|---|---|
| 司法試験 短答式 2026, points /175 (pass 99) | 130 (vote@4) | 101 | 92 | 82 |
| 司法試験 短答式 2024, points /175 (pass 93) | 129 (vote@4) | 90 | 100 | 96 |
| 司法試験 短答式 2025, points /175 (pass 81) | 120 (vote@4) | 87 | 88 | 85 (fail) |
| Benrishi-MC (弁理士 短答 2024-25), accuracy | 0.517 (vote@4) | 0.350 | 0.383 | 0.383 |
| LawQA-JP, accuracy | 0.821 (vote@4) | 0.821 | 0.793 | 0.800 |
| JBE-QA (200-item sample), MCC | 0.537 | 0.534 | 0.600 | 0.660 |
| OI-Contract Review, checklist recall (with adapter) | 0.240 | 0.228 | 0.169 | 0.173 |
| OI-Contract Draft, row score (with adapter) | 0.829 | 0.794 | 0.759 | 0.692 |
Single runs (no voting) average 115.8 (2026) and 115.0 (2024). Claude Opus 5.5, closed-book, scores 171 and 173; on contract review it reaches 0.475.
The scores above depend on the retrieval context: without it the base model scores 20-33 points lower per exam run.
serve/ and web/ in the code repository run the full stack (retrieval gateway, contract adapter, voting):
bash
bash serve/start_vllm.sh /path/to/weights /path/to/contract-adapter # vLLM OpenAI server, FP8, 48 GB GPUpython -m uvicorn serve.gateway:app --port 8080 # retrieval + voting gatewayDecoding: thinking off (enable_thinking=False), temperature 0.7, top-p 0.8, top-k 20, min-p 0, presence penalty
1.5; 8,192 new tokens for exam answers. Greedy decoding loops on long answers.
Government sources (法務省, 特許庁, 農林水産省, 経済産業省, 国土交通省, e-Gov) are used under 公共データ利用規約 第1.0版
(PDL1.0): 出典を明記し、編集・加工して利用しています。See docs/release/LICENSE_NOTES.md for the full list.
bibtex
@misc{shin2026xlawjp, title = {XLaw-JP-27B: A Lightweight Open Model Specialized for Japanese Law}, author = {Shin, Andrew}, year = {2026}, eprint = {2610.39227}, archivePrefix = {arXiv}, primaryClass = {cs.CL}, url = {https://arxiv.org/abs/2610.39227}}Apache 2.0, as the base model. These weights are a modified version of
Qwen/Qwen3.8-27B (Copyright Alibaba Cloud / Qwen Team, Apache 2.0),
fine-tuned by X-Law, Inc.; the base model's licence is included as LICENSE. Attribution for the government sources used in training is listed in
docs/release/LICENSE_NOTES.md (出典の明記、編集・加工した旨).
bash serve/start_vllm.sh /path/to/weights /path/to/contract-adapter # vLLM OpenAI server, FP8, 48 GB GPUpython -m uvicorn serve.gateway:app --port 8080 # retrieval + voting gateway@misc{shin2026xlawjp, title = {XLaw-JP-27B: A Lightweight Open Model Specialized for Japanese Law}, author = {Shin, Andrew}, year = {2026}, eprint = {2610.39227}, archivePrefix = {arXiv}, primaryClass = {cs.CL}, url = {https://arxiv.org/abs/2610.39227}}