Modelpublic

websfactory/Webs-Sejong-31B-v7

sourceHugging Facegemmaupdated 3mo agoView on Hugging Face

1likes8downloads

Model Card

Webs-Sejong-31B

🏆 Ranked #1 on the K-AI Leaderboard (leaderboard.aihub.or.kr) Leaderboard entry: Webs-Sejong-31B-R1 Public release: websfactory/Webs-Sejong-31B-v7 Overall average 0.624, ranked #1 as of 2026-07-05. Evaluated on the K-AI Leaderboard, a public Korean LLM evaluation platform operated through AI Hub / NIA, using non-public benchmark data that is not disclosed to participants.

Webs-Sejong-31B is a 31B-parameter Korean-centric language model based on google/gemma-4-31B-it. It is strong at Korean-language knowledge, Korean cultural context, professional and academic reasoning, and commonsense QA, while retaining English capability. This repository provides the same checkpoint that was submitted as Webs-Sejong-31B-R1 on the K-AI Leaderboard.

Highlights

—#1 on the K-AI Leaderboard. Overall average 0.624, the top score on the public board as of 2026-07-05 (leaderboard entry: Webs-Sejong-31B-R1).
—Korean-first. Strong on Korean cultural and academic tasks, with English ability retained.
—Drop-in Gemma-4. Standard Gemma-4 architecture and tokenizer: compatible with the Hugging Face transformers Gemma-4 implementation and expected to work with Gemma-4-compatible serving stacks.

Evaluation: K-AI Leaderboard

Evaluated on the K-AI Leaderboard, a public Korean LLM evaluation platform operated through AI Hub / NIA. Scores are produced on non-public benchmark data that is not disclosed to participants.


Leaderboard entry	`Webs-Sejong-31B-R1`
Overall average	0.624
Rank	#1 (as of 2026-07-05)

Because the benchmark data is not disclosed to participants, this reduces the likelihood of direct benchmark overfitting. Users should still evaluate the model on their own target tasks.

Model


Architecture	Gemma-4-31B (dense)
Parameters	~31B
Precision	bfloat16
Languages	Korean (primary), English
Base model	`google/gemma-4-31B-it`

Hardware

At bf16 the weights are roughly 62 GB. Practical setups:

—Full precision: one 80 GB GPU (A100 / H100), or two 40–48 GB GPUs.
—4-bit quantized: roughly 20–24 GB for the quantized weights; allow extra memory for KV cache, context length, batch size, and image inputs.

Usage

python

from transformers import AutoModelForImageTextToText, AutoProcessor

model_id = "websfactory/Webs-Sejong-31B-v7"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(model_id, device_map="auto")

This model follows the standard Gemma-4 interface. For image-and-text input formatting, refer to the base model documentation at `google/gemma-4-31B-it`.

Training Details

Training and adaptation details are proprietary and are not disclosed in this release.

Intended Use & Limitations

Intended for Korean-language assistance, knowledge QA, and reasoning. Like any language model it can produce incorrect or outdated information, so do not rely on it for medical, legal, financial, or public-policy decisions without human review. Evaluate it on your own target tasks before production deployment.

License

This model is a derivative of Gemma-4 and is distributed under the **Gemma Terms of Use**. By using this model you agree to those terms and Google's Prohibited Use Policy.