LGAI-EXAONE

/

K-EXAONE-236B-A23B

like 558

Follow

LG AI Research 1.66k

Dear K-EXAONE team,

First of all, thank you for making K-EXAONE available to the community. Your work on building a world-class Korean-native LLM is truly inspiring, and it has been a privilege to build on top of it.

I am a computer science undergraduate student in South Korea. I am currently developing KOSMOS, an open-source conversational AI platform that aims to turn Korea's 5,000+ fragmented government APIs (data.go.kr) into a single natural-language interface for citizens.

K-EXAONE (236B-A23B) serves as the core LLM for this platform, accessed through the FriendliAI Serverless API. I am reaching out to humbly ask whether there might be any form of academic credit or compute support available for student researchers using K-EXAONE in public-interest projects. I completely understand if this is not something that can be accommodated, and I appreciate you even taking the time to read this.

Why K-EXAONE

I would like to share why K-EXAONE was the natural and only choice for KOSMOS:

1. Native Korean language capability — Government APIs return Korean-only responses with inconsistent schemas across 5,000+ endpoints. K-EXAONE's exceptional Korean proficiency is essential for accurate intent analysis and citizen-facing answers. No other model matches this level of Korean understanding.
2. MoE cost efficiency — The 236B total / 23B active parameter architecture delivers strong reasoning at a fraction of dense-model inference cost. For a self-funded student project calling LLMs in a multi-turn tool loop, this efficiency makes the project feasible at all.
3. Tool-use competency — The platform runs an async tool loop where the LLM must reliably select, parameterize, and chain public API calls across government ministries. K-EXAONE has performed remarkably well in this demanding scenario.
4. Alignment with Korean public AI goals — Korea's AI Action Plan (2026-2028) emphasizes public-sector AI transformation. Building this platform on a Korean-made LLM feels right — KOSMOS validates principles 8 and 9 of that national plan.

About the Project

The platform is under active development with a spec-driven workflow, and all source code is publicly available for review.

A Humble Request

As a university student, I am currently self-funding all API costs. FriendliAI Serverless credits for K-EXAONE inference are the primary expense, and as the project grows, sustaining this is becoming a challenge. If any of the following were possible, I would be truly grateful:

• Academic credit allocation for continued development and experimentation
• Guidance on any existing student or researcher support programs I may not be aware of
• A referral to the appropriate contact if this is not the right channel

In return, I would be more than happy to provide a detailed technical report on K-EXAONE's performance in the Korean public-service domain, share benchmark results, or properly acknowledge LG AI Research's contribution in any resulting publications.

What This Project Could Offer Back

I hope this project can be a small contribution back to the K-EXAONE ecosystem as a real-world validation in a meaningful domain:

• Multi-turn Korean tool-use across heterogeneous government APIs
• Permission-gated PII handling under Korea's PIPA (Personal Information Protection Act)
• Cost-optimized prompt caching in long citizen sessions (128K context window)
• MoE routing behavior analysis under real Korean public-service workloads

All findings will be shared openly with the community.

Thank you so much for your time, and once again, thank you for building K-EXAONE. It is genuinely enabling work that would not be possible otherwise.

With gratitude,
Um Yunsang
Computer Science, Undergraduate
GitHub: github.com/umyunsang