이거 모르고 장비 사면 돈 날립니다 | 로컬 LLM 현실 + 구축 가이드 (AMD R9700 + Hermes)

Without cloud subscription fees, token fees, or worries about data leaks, we will run an AI agent entirely locally on a single workstation PC. "Local LLM, where do I start?" Laptops and mini-PCs only offer a taste of it; To run it properly, you ultimately need a dedicated PC. To help beginners avoid getting lost, this video honestly outlines, based on actual measurements of the AMD Radeon AI PRO R9700 32GB, ① which equipment to choose, ② what criteria to use to select a model, and ③ how to build a budget. Then, by layering the Hermes agent on top of that, we will create your own local assistant that handles everything from web searches to file creation and editing on its own. ▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬ 🖥️ Today's Equipment for Actual Measurement (Sponsored by AMD) ▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬ • CPU: AMD EPYC 4585PX (16 Cores · 3D V-Cache) • GPU: AMD Radeon AI PRO R9700 32GB (RDNA4 · Navi 48) • "The cheapest 32GB card capable of running the best models right now" — Price compared to 5090, VRAM compared to 5080 🛒 EPYC 4585PX Purchase/Special Offer: https://www.compuzone.co.kr/product/p... ▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬ ⚡ AMD AI PC Launcher — One-Click Setup ▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬ If you find it burdensome to set it up manually like in the video, there is a one-click launcher for AMD systems. With just one button, it automatically handles everything from downloading local models → connecting the Hermes agent → web search, documentation tools, and Telegram bot integration. It is an especially easy starting point for Windows users. 🔗 GitHub: https://github.com/iamchobosalsal/ai_... ▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬ ⏱️ Timestamp ▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬ 00:00 Intro 02:06 Today's Guide Step 3 03:27 Meaning of Parameter 'B' 04:24 VRAM Tier Chart 06:33 Price Reality & Equipment Specs 12:19 Quantization (Q4) and Required VRAM Formula 13:17 Finding the Right Model for My Equipment (which-llm) 16:29 Understanding Token, MoE, and Dense Concepts 19:23 Model Selection Order & Context (KV Cache) 20:51 Structure of Speed ​​(Prefill, TTFT, Decode) & 5-Axis Measurement 23:06 Performance Cliff Section & Korean Language & License Summary 26:10 LM Studio Introduction, Installation & AMD ROCm 31:39 Model Download & Load & Launching Inference Server 33:20 LM Link — Remote Use on Laptop & Mobile 41:48 Hermes Agent Integration (Custom Endpoint) 48:15 Preparing for the Real Battle — Models to Compete Today 53:32 Gemma3 30B 55:02 Qwen3 30B (MoE) 61:05 Qwen3.6 27B (Dense) 67:17 Qwen3.6 35B (MoE) 71:44 GLM 4.7 · Kanana 2 · Exa One 77:51 GPT-OSS 20B 80:28 Conclusion — What Works vs. What Doesn't 83:06 Outro — Budget Baseline & AMD AI PC Launcher ▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬ Local LLM tested in the video (All Q4) ▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬ • gpt-oss-20b (OpenAI · MoE 20B) — #1 in Speed ​​· Agent Finish Balance, the mainstay of this gear https://huggingface.co/openai/gpt-oss... • Qwen3-30B-A3B (Alibaba · MoE 30B) — MoE proving the value of 32GB, Apache 2.0 https://huggingface.co/Qwen/Qwen3-30B... • Qwen3.6-27B (Alibaba · Dense 27B) — Latest generation Dense flagship, Apache 2.0 https://huggingface.co/Qwen/Qwen3.6-27B • Qwen3.6-35B-A3B (Alibaba · MoE 35B) — Bigger and faster MoE, Apache 2.0 https://huggingface.co/Qwen/Qwen3.6-3... • kanana-2-30b-a3b (Kakao · MoE 30B) — Domestic Korean agent Leading Stocks https://huggingface.co/kakaocorp/kana... • EXAONE-4.0-32B (LG AI · Dense 32B) — Ranked 1st in Actual Korean Expression Quality https://huggingface.co/LGAI-EXAONE/EX... • Gemma 4 31B (Google · Dense 31B) — Highly Interested New Flagship https://huggingface.co/google/gemma-4... • GLM-4.7-Flash (z.ai · MoE 30B) — A sensational title claiming to be "the best in the 30B class," MIT https://huggingface.co/zai-org/GLM-4.... ※ Please be sure to check the license on each model card before commercial use (Note the non-commercial conditions for kanana and EXAONE). ▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬ Inflearn Course — Hermes × Codex ▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬ If you would like to delve deeper into the Hermes agent integrated into a local LLM as seen in this video? You can find the entire process in this Inflearn course to directly hire AI employees for your own virtual company using multi-agents and build an automated company that operates 24/7. 🎓 Inflearn Course (Hermes × Codex) : https://inf.run/GodUp ▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬ 🔧 Tools & Links Covered in the Video ▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬ • LM Studio — Local LLM Executor (Headless CLI · LM Link Remote) https://lmstudio.ai • Hermes Agent (Nous Research) — AI Agent deployed locally https://hermes-agent.nousresearch.com • whichllm — A library that recommends models suitable for your hardware based on benchmarks https://github.com/Andyyyy64/whichllm • AMD ROCm — AI compute stack for AMD GPUs (compatible with NVIDIA CUDA) https://rocm.docs.amd.com • Hugging Face — Distribution site for openweight models (download the...