inflearn logo

Complete Analysis of the Solar Open 2 Architecture

An AI that understands Korean well, handles long documents, and even carries out real-world tasks—that is exactly the direction Solar Open 2 is aiming for. This class provides an accessible explanation of Upstage’s Solar Open 2 technical report in the order of “Introduction → Design → Training → Evaluation.” Technical terms are explained through everyday analogies and visual materials, enabling even beginners to understand “what,” “how,” and “why” as one coherent flow.

15 learners are taking this course

Level Beginner

Course period Unlimited

AI
AI
LLM
LLM
AI Agent
AI Agent
AX(Agent Experience)
AX(Agent Experience)
foundation
foundation
AI
AI
LLM
LLM
AI Agent
AI Agent
AX(Agent Experience)
AX(Agent Experience)
foundation
foundation

What you will gain after the course

  • Understand the essentials of the Solar Open 2 technical report (introduction → design → training → evaluation) so that even beginners can follow along.

  • You will be able to explain terms such as MoE, hybrid attention, NoPE, pre- and post-training, agents, and Ko-GDPval in context.

A Complete Analysis of the Solar Open 2 Architecture

Read Technical Reports Easily

Do you only know Solar Open 2 by name?

This course is an easy-to-follow class that explains Upstage’s publicly released Solar Open 2 technical report in a way that even beginners can understand.

Korean AI has moved beyond merely “translating well” and entered a stage where it can read long documents and even carry out real-world tasks. Solar Open 2 is one of the key models at the center of this shift. However, the original technical report is dense with terminology, and its architecture, training, and evaluation are presented all at once, making it difficult to read on your own.

So this course selects only the key points from the report and

We explain it in the order of introduction → design → training → evaluation.

What 250B in scale, 15B active parameters, and a 1M-token context mean

Why designs such as MoE, hybrid attention, and NoPE are necessary

How to build agents through pretraining and post-training

What the English and Korean scorecards, Ko-GDPval, and real-world work cases tell us

We first establish technical terms through everyday analogies and visual materials, then reinforce them with the necessary figures and experimental results.

This course focuses on understanding, interpretation, and the ability to explain rather than coding exercises.

Recommended for AI planners, developers, students, and anyone who wants to properly assess the capabilities of Korean-language AI.

After completing the course, you will understand Solar Open 2 not merely as a model you know by name, but well enough to explain its design intent and strengths.

Structure: 5 chapters · 15 lessons · Approximately 1 hour 49 minutes in total


What you’ll learn

1⃣Understand at a glance what Solar Open 2 is and MoE

We cover the model specifications (scale, active parameters, and context), the need for Korean-specific AI, comparisons between MoE and generations, the report card, and the overall roadmap.

2⃣Model Blueprint — Hybrid Attention, NoPE, and 1M Tokens 

We delve into the design changes that made long texts easier to read. Follow the combined effects of hybrid attention, NoPE and auxiliary mechanisms, and expert expansion.

3⃣ From training to evaluation — agents and real-world report cards

Learn how to build daily-life, coding, and office agents through pre- and post-course study, and how to read their report cards using English and Korean benchmarks, Ko-GDPval, and real-world work cases.

Notes Before Taking the Course

Learning materials

Prerequisites and Notes

  • No prior knowledge is required.

  • It will be easier if you have a general sense that “an LLM reads tokens and predicts what comes next,” or a casual interest in AI and open models.

  • The basic concepts needed for the course will be explained as they come up, so all you need to do is follow along enthusiastically.

  • This course focuses on explaining the Solar Open 2 technical report and does not include writing code for model training or deployment practice.

  • The copyright for the learning materials and lecture content belongs to the creator. Unauthorized distribution, reproduction, and use for purposes other than personal study are prohibited.

Recommended for
these people

Who is this course right for?

  • For those who want to quickly get up to speed on the trends in Korean AI and open models

  • For those who are unfamiliar with the terms MoE, attention, and agents but want to understand them

  • Planners, developers, and students who find technical reports overwhelming

Need to know before starting?

  • A basic understanding that an LLM “reads a sentence in token units and predicts what comes next”

  • Experience having heard terms like parameters, benchmarks, and agents at least once.

  • (Nice to have) A casual interest in Korean AI and open models

Hello
This is DEVJH

94

Learners

6

Reviews

5.0

Rating

5

Courses

Hello. I’m Jonghyun Lee.

While working as a software developer at Coupang, I gained hands-on experience developing and operating real-world services. I am currently balancing work and studies while researching machine learning and deep learning in the Master’s program in Artificial Intelligence at Yonsei University.

With over 10 years of software development experience, I conduct AI research firsthand and implement it through various projects and services. I am particularly interested in machine learning and deep learning, LLMs and generative AI, RAG, AI agents, and AutoML. Rather than merely studying these technologies, I value the process of implementing and experimenting with them directly to solve problems.

Based on hands-on experience developing AI models, I have participated in various AI competitions and achieved strong results. In the 2026 K-League–University of Seoul Open AI Competition, I placed third out of 947 teams, and in the 5th ETRI Artificial Intelligence Paper Competition on Human Understanding, I ranked second on the Private Leaderboard out of 409 teams.

I have also continued to achieve results in research. My work on FS-DCM, which addresses the nonstationarity and volatility of time-series data, was accepted as a first-author oral presentation paper at the International Conference on Automated Machine Learning (AutoML 2026), while my first-author paper, MathQueue, was also accepted at the EMNLP 2026 MathNLP Workshop. In addition, I continue to pursue research on interpretable AutoML using LLMs and the practical application of AI technologies.

I currently operate the AI learning service Commuting AI (https://mdooai.com). I also work as a Upstage Certified Trainer (Upstage-certified AI trainer), providing generative AI education and mentoring.

Based on my practical experience in development, AI research, and education, I work with everyone from those just starting out in AI to those looking to expand their development experience into the AI field, helping them consider realistic learning paths and project experiences suited to their individual circumstances.

Going forward, I aim to contribute to helping more people understand AI, implement it themselves, and apply it effectively to their work and projects, based on AI technologies and research that I have directly experienced and validated.

Thank you.

More

Curriculum

All

15 lectures ∙ (1hr 49min)

Published: 
Last updated: 

Reviews

Not enough reviews.
Please write a valuable review that helps everyone!

DEVJH's other courses

Check out other courses by the instructor!

Similar courses

Explore other courses in the same field!

25% off for new members

$8.80

25%

$8.80