RableUp Unveils Korean-Specialized “LLM Draft Model” as Open Source
IT DAILY · · 1 views
✦ AI Summary
RableUp announced on the 12th that it will release the Korean-specialized draft model “GLM-5.2-NVFP4-DFlash-ko” as open source.
The model is a draft model that applies z-lab’s “DFlash” technique, and runs in NVIDIA NVFP4 format when combined with Z.ai’s open LLM “GLM-5.2.”
In actual Korean agent traffic environments, the average acceptance length (MAL) was 2.71 for agent tasks and 2.77 for Korean queries, and RableUp is also planning to develop a Korean draft model for Z.ai’s “GLM-5.3-Flash” and independently train a Korean diffusion language model for on-device agents.
RableUp announced on the 12th that it will release the Korean-specialized draft model “GLM-5.2-NVFP4-DFlash-ko” as open source. The model is intended to improve LLM inference speed for Korean queries and agent tasks.
“GLM-5.2-NVFP4-DFlash-ko” is a draft model that applies z-lab’s “DFlash” technique. The model is designed to be combined with Z.ai’s open LLM “GLM-5.2” and runs in NVIDIA NVFP4 format. It is distributed through RableUp’s repository on Hugging Face.
The draft model is used for “Speculative Decoding.” In this approach, a small model pre-predicts multiple next tokens, and the main large model verifies the predictions in batches. The defining feature of the DFlash approach is that it generates tokens in parallel in batches rather than sequentially, thereby reducing inference time.
Existing draft models built on English data had low prediction hit rates in Korean environments, which in turn reduced acceleration efficiency. RableUp conducted Korean-language data training to address this issue.
RableUp improved the existing problem through training on Korean data. In actual Korean agent traffic environments, the model’s average acceptance length (MAL) was 2.71 for agent tasks and 2.77 for Korean queries.
The model’s performance is on par with English DFlash models. By contrast, the MAL of English-trained draft models was in the 1.x range.
RableUp plans to develop a Korean draft model for Z.ai’s “GLM-5.3-Flash.” It is also independently training a Korean diffusion language model for on-device agents that can run on local terminals, with a model size of 11 billion parameters. For this training, it used support from the Ministry of Science and ICT and the National IT Industry Promotion Agency (NIPA) under the “Advanced GPU Utilization Support Project,” and deployed a GPU cluster with 64 NVIDIA B200 units.
RableUp CEO Shin Jeong-gyu said existing draft models were trained mainly on English, so the acceleration effect of Speculative Decoding was limited in Korean-language environments. Accordingly, RableUp said it released the model with the goal of securing acceleration performance in Korean environments for AI service developers, and added that it is expanding its capabilities through in-house model development based on its infrastructure software expertise.
Meanwhile, RableUp has submitted a securities registration statement for its Kosdaq listing. It will also conduct book-building for institutional investors from the 13th to the 19th.
Source: IT DAILY · Kwon Young-seok
Original: https://www.itdaily.kr/news/articleView.html?idxno=242119
References
This article was produced with the help of an automated content generation algorithm.
Source: IT DAILY
View originalThis article was summarized and organized by BizCrush based on the original article from IT DAILY. For exact quotations and full details, please refer to the original article.