llm-jp-4.1-33b-thinking

LLM-jp-4.1 is a series of large language models developed by the Research and Development Center for Large Language Models at the National Institute of Informatics.

This repository provides the llm-jp-4.1-33b-thinking model. For an overview of the LLM-jp-4.1 models across different parameter sizes, please refer to:

• LLM-jp-4.1 Models

Base models are trained with pre-training and mid-training only. Post-trained models are aligned using supervised fine-tuning (SFT) and direct preference optimization (DPO), without reinforcement learning.

For more details on the training procedures and evaluation results, please refer to our technical blog (in Japanese).

For practical usage examples and detailed instructions on how to use the models, please also refer to our cookbook.

To support the continued development of LLM-jp, we would greatly appreciate it if you could share how you utilize LLM-jp outcomes via the survey form.

Usage

Please refer to our cookbook for practical usage examples and detailed instructions on how to use the models.

Model Details

• Model type: Transformer-based Language Model
• Architectures:

Dense model:

MoE model:

Tokenizer

The tokenizer of this model is based on a Unigram byte-fallback model implemented with huggingface/tokenizers. The vocabulary entries were converted from llm-jp-tokenizer v4.0. Please refer to README.md of llm-jp-tokenizer for details on the vocabulary construction procedure (the pure SentencePiece training does not reproduce our vocabulary).

[!NOTE] The chat template of this model is designed to be compatible with the OpenAI Harmony response format. However, the tokenizer differs from the one assumed by the openai-harmony library, and therefore direct tokenization with openai-harmony is not supported. For correct behavior, please use the tokenizer provided with this model. For detailed usage, please refer to our cookbook.

Training

Pre-training

This model was trained through a multi-stage pipeline consisting of pre-training and mid-training phases, using a total of 11.7T tokens.

The corpora used for pre-training and mid-training are publicly available at the following links:

• Pre-training
• Mid-training

[!NOTE] Although most of the corpora have been released, some portions are excluded from public release due to licensing constraints.

Post-training

We have fine-tuned the pre-trained checkpoint using SFT and further aligned it with DPO.

The datasets used for post-training are also publicly available at the following links:

• SFT
• DPO (for llm-jp-4.1-8b-thinking model)
• DPO (for llm-jp-4.1-32b-a3b-thinking model)
• DPO (for llm-jp-4.1-33b-thinking model)

Evaluation

We evaluated llm-jp-4.1 on a variety of benchmarks covering general capabilities, safety, and tool calling.

For more detailed evaluation results and analysis, please refer to our technical blog.

swallow-evaluation-instruct

We evaluated the models on a range of benchmarks covering the following six categories:

• Math
  • Math 500
  • AIME 2024 (pass@1, pass@32)
  • AIME 2025 (pass@1, pass@32)
  • AIME 2026 (pass@1, pass@32)
  • MCLM Math 100 (pass@1, pass@4)
  • PolyMath JA High
  • PolyMath JA Top
• Science
  • GPQA Diamond (pass@1, pass@4)
  • JGPQA Diamond
• Knowledge & QA
  • JAM-CQA
  • JEMHopQA
  • JMMLU
  • MMLU-ProX JA
  • MMLU-ProX EN
• Code
  • LiveCodeBench v6 (pass@1, pass@10)
  • JHumanEval (pass@1, pass@10)
  • HumanEval+ (pass@1, pass@10)
• Instruction Following (IF)
  • MIFEval JA
  • IFBench
• Machine Translation (MT)
  • WMT20 EN-JA
  • WMT20 JA-EN

For LLM-jp and gpt-oss models, reasoning_effort was set to high. For Olmo-3-7B-Think, Olmo-3.1-32B-Think, Qwen3, Qwen3.5, Qwen3.6, and Gemma 4, enable_thinking was set to True. For Qwen3.8-27B and Muse-Glimmer-30B, reasoning_effort was set to xhigh.

The figure below shows the average score across the benchmarks in each category.

llm-jp-judge

We evaluated the models using an LLM-as-a-Judge framework on the following benchmarks:

• MT-Bench (JA/EN): A benchmark for measuring multi-turn conversational task-solving ability.
• AnswerCarefully: A benchmark for evaluating safety in Japanese. We used 336 questions from the v2.0 test set.
• llm-jp-instructions: A set of human-created single-turn question-answer pairs. We used 400 questions from the test set.

We used gpt-5.4-2026-03-05 as the judge. For models that support reasoning_effort, it was set to medium.

The scores represent the average values obtained from three rounds of inference and evaluation. For more details, please refer to the evaluation code.

Tool Calling

We evaluated the models on the following tool calling benchmarks:

• BFCL
• tau2-bench

For BFCL, we evaluated the models using only the categories available up to v3, so the web search and memory categories introduced in v4 are excluded. For tau2-bench, we used Azure's gpt-5.1-2025-11-13 as both the user simulator and the NL-assertion judge, and the scores represent the average values obtained from two rounds of inference and evaluation. For all LLM-jp models, reasoning_effort was set to medium.

The figure below shows the scores on each benchmark.

Risks and Limitations

The models released here are research and development models and are not intended for direct use in production services. Although the models have undergone post-training for instruction following and safety, they may still generate inaccurate, inappropriate, or otherwise undesirable outputs. Users should carefully evaluate the models for their intended use cases.

Send Questions to

llm-jp(at)nii.ac.jp

License

Apache License, Version 2.0

Acknowledgements

To develop this model, we used the NINJAL Web Japanese Corpus (whole-NWJC) from the National Institute for Japanese Language and Linguistics (NINJAL).

Model Card Authors

The names are listed in alphabetical order.

Hirokazu Kiyomaru, Takashi Kodama, and Yunang Wu.