llm-jp-4.1-33b-thinking

llm-jp-4.1-33b-thinking

A 33.2B-parameter English and Japanese text generation model from the LLM-jp project. Post-trained with SFT and DPO, it has a 65,536-token context, Apache 2.0 weights and loads through transformers.

Made by llm-jp on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: llm-jp-4.1-33b-thinking is a 33.2B-parameter text generation model from the LLM-jp project. The project comes from a language model research center at the National Institute of Informatics. This is the post-trained 33B size of the LLM-jp-4.1 series, for English and Japanese.

It went through pre-training and mid-training on 11.7T tokens in several stages. The team then tuned it with supervised fine-tuning and aligned it with direct preference optimization, with no reinforcement learning. The card says the post-training datasets are public.

Who it suits: Developers and researchers who want an open Japanese and English model they can host themselves, and teams who study how a model built on public corpora is trained.

Industry | Who uses it | How they use it | Example
Research | NLP researchers | Study open training data and recipes | Comparing the 33B size with the 8B size
Software | Application developers | Host a Japanese and English assistant | A support bot that answers in both languages
Education | Language teams | Build tools around Japanese text | A helper that explains a Japanese passage
Enterprise | Platform teams | Run a model on their own servers | Internal question answering with no outside service

You can skip it if you only have a small GPU, since the weights are large, or if you need languages beyond English and Japanese.

Where to use it: It loads through the transformers library in Python. The card sends you to the LLM-jp cookbook for usage examples and detailed steps.

• Python with transformers
• Windows, macOS or Linux, per the Hugging Face steps
• A host with enough memory for the weights
• The tokenizer shipped with the model

Use case | Who | What they get back
Chat answers | App developers | Generated text in English or Japanese
Long prompts | Data teams | Input up to a 65,536-token context
Model study | Researchers | An open recipe to compare with other sizes

How to set it up: You need Python and a machine that can hold a 33B model. The card page lists about 66.4 GB of weights in BF16 and about 16.6 GB at 4-bit, as an estimate.

• Python with pip and transformers
• About 66.4 GB of disk for the BF16 weights
• A GPU with a lot of memory, or a quantized copy from another author

1. Install the Hugging Face hub tools with pip.
2. Download the model files with the hf command.
3. Follow the LLM-jp cookbook to load the model with transformers.

pip install -U huggingface_hub
hf download llm-jp/llm-jp-4.1-33b-thinking

How to use it: Load the model with its own tokenizer and chat template, send a conversation and read the reply. The card shows no code here and points to the cookbook for exact steps.

The chat template is built to work with the OpenAI Harmony format, but the tokenizer is different. The card says direct tokenization with the openai-harmony library is not supported, so use the tokenizer that comes with the model.

Example: You load the model with its tokenizer and ask a question in Japanese. You then ask the same question in English and compare the two replies.

For newcomers and experts: If you are new, read the cookbook first and try one short prompt before you send long ones. Check memory use early, because a 33B model is large.

If you know the field, this dense model has 64 layers, a hidden size of 5,120, 40 heads and a 65,536-token context. The tokenizer is a Unigram byte-fallback model, with vocabulary converted from llm-jp-tokenizer v4.0.

Prompt in English or Japanese
Chat template and model tokenizer
Model generates a reply
Text out

Good to know: The card lists dense 8B and 33B models plus a 32B-A3B mixture-of-experts model in the series. Some training corpora are not public because of licensing limits. The chart shows the two weight sizes the page lists.

Title: Weight size, GB
BF16: 66.4 GB
4-bit estimate: 16.6 GB

• License: Apache 2.0
• Cost to run: none named, and the weights need a lot of memory
• Data it reaches: the prompts you send, and the model files you download
• Not for: languages other than English and Japanese, since the card lists only those two

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 33.2B
• Weights: about 66.4 GB in BF16; about 16.6 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 4,224
• Likes: 17
• License: apache-2.0
• Languages: en, ja
• Quantized versions: 7
• Fine-tunes: 2
• Tags: transformers, safetensors, llama, text-generation, conversational

Links

• Model: https://huggingface.co/llm-jp/llm-jp-4.1-33b-thinking
• Files: https://huggingface.co/llm-jp/llm-jp-4.1-33b-thinking/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download llm-jp/llm-jp-4.1-33b-thinking

Then follow the usage in its model card. Nothing in this kit runs by itself.