Kakao Brain (KoGPT)
Kakao Brain's 2021 Korean GPT, kept online as a non-commercial research artefact
Overview
KoGPT is a 6.17-billion-parameter Korean generative language model released by Kakao Brain in November 2021. Its weights remain downloadable from Hugging Face for non-commercial research under CC BY-NC-ND 4.0, with Apache 2.0 code on GitHub, but its API closed on 30 September 2024 and Kakao Brain's AI team was absorbed into Kakao. Kakao's current models are the Kanana family.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 14 Oct 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
Use cases
What you can produce with Kakao Brain (KoGPT)
- KoGPT 6.17B weights in float32 and float16 revisions, downloadable from Hugging Face for research
- Apache 2.0 inference code on GitHub, compatible with Hugging Face Transformers
- A 64,512-token tokeniser built for Korean text
- Few-shot Korean text classification and generation for reproducing 2021-era benchmarks
- A documented baseline against HyperCLOVA on NSMC, YNAT and KLUE-STS
ASEAN Perspective
Kakao Brain (KoGPT) in Southeast Asia
KoGPT is a Korean-first model with limited direct utility for ASEAN languages — its vocabulary and training data are almost exclusively Hangul, and the planned multilingual KoGPT 2.0 (which would have added Bahasa and Vietnamese support) was never released. That said, the Korean diaspora and Korean-content consumption across the Philippines, Indonesia, Malaysia, and Vietnam (driven by K-pop and K-drama) creates genuine demand for Korean NLP tools in the region. For ASEAN-based researchers studying Korean or building Korean-language consumer apps, KoGPT remains a freely accessible, academically credible baseline — though teams needing full ASEAN language coverage will need to supplement it with multilingual models.
KoGPT mattered in 2021 as an openly downloadable Korean model at six-billion-parameter scale, and its few-shot results against HyperCLOVA were respectable. Today it is a legacy artefact: no API since September 2024, no updates since January 2024, a successor that never shipped, and a licence that bars commercial use and modified redistribution.
Download it only to reproduce older Korean NLP research. For anything current, Kakao's own Kanana models, LG's EXAONE or Naver's HyperCLOVA X are the places to look.
What we found
KoGPT is a Korean generative language model that Kakao Brain, then Kakao's AI research subsidiary, released in November 2021. It has 6.17 billion parameters and a 64,512-token vocabulary built for Hangul, and was trained on an in-house Korean corpus Kakao Brain called "ryan". The model card reports few-shot scores of 87.8% accuracy on NSMC sentiment classification and 78.0 F1 on YNAT topic classification, set against Naver's HyperCLOVA at several sizes. Those are Kakao Brain's own figures, and its fine-tuning results were withdrawn from the README after a GitHub issue questioned whether the comparison was fair.
Everything around the model has since been wound down. Kakao absorbed Kakao Brain's AI development team in June 2024, ended the KoGPT and Karlo image-generation APIs on 30 September 2024, and merged what remained of Kakao Brain with its IT subsidiary dK Techin from 1 October 2024. The kakaobrain.com website no longer resolves. KoGPT 2.0, promised in sizes up to 65 billion parameters for late 2023, never shipped, and the repository has not changed since January 2024.
What is left is a research artefact. The weights can still be downloaded from Hugging Face after accepting the access conditions, and the inference code on GitHub is Apache 2.0. The weights themselves are CC BY-NC-ND 4.0, which forbids commercial use and the distribution of modified versions, and the card asks for 32GB of GPU memory, or 16GB at half precision. Hugging Face recorded ten downloads in the past month when checked in October 2026.
Kakao's current work carries other names. It partnered with OpenAI in 2025 and launched ChatGPT for Kakao inside KakaoTalk that October, and it publishes its own Kanana models, including 1.3B and 3B Kanana-2 releases in July 2026. The standalone Kanana app closes on 15 October 2026. Kanana has its own entry in this directory; KoGPT is worth downloading only to reproduce older Korean NLP papers.
Compiled by RECATOOLS from public sources: vendor documentation and pricing, code repositories and published reporting.
Notable facts
- KoGPT was trained on a corpus called the 'ryan' dataset — named internally at Kakao Brain — and uses a custom 64,512-token vocabulary optimised for Hangul morphology.
- The model requires a minimum 16GB GPU (float16) or 32GB GPU (float32) to run locally — making it one of the more hardware-demanding open research models of its era.
- Kakao founder Kim Beom-su personally served as the first CEO of Kakao Brain when it was spun out in February 2017 with 20 billion KRW (~$17.5M) in starting capital.
- Despite planning a KoGPT 2.0 with up to 65 billion parameters for Q3 2023, Kakao eventually pivoted to an OpenAI partnership — embedding GPT-5 into KakaoTalk, which hit 8 million users within three months of launch.
Frequently asked questions
About this listing
This entry was compiled from publicly available data including Kakao Brain (KoGPT)'s official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with Kakao Brain (KoGPT) unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to Kakao Brain (KoGPT) directly →
Spotted something out of date? Suggest an update →
Alternatives to Kakao Brain (KoGPT)
More in LLMs & Chat
