JP | EN

Exploring Local LLM Workflows

A Practical Digest of Tools, Models, and Use Cases

Local AI Workstation Setup

On this page:

With recent improvements in consumer GPUs, tooling, and open-weight models, running large language models locally has become not only feasible but genuinely useful. I set up my PC as a local AI workstation and tested several real-world LLM-related use cases, focusing on productivity, automation, multimodal generation, and developer workflows.

This post is a high-level digest of what I tried: the applications, models, use cases, and required software. Each topic will be expanded into a separate, detailed post covering installation, configuration, and concrete generation examples.

Base Environment Overview

Before diving into individual applications, here is the PC specs and foundational software stack used across all experiments.

Part Type Product Name Manufacturer Main Specifications
CPU Core Ultra 7 265K Intel Arrow Lake-S architecture, unlocked processor, high-performance desktop CPU
CPU Cooler Peerless Assassin 120 Black Thermalright Dual-tower air cooler, 120 mm fan, 6 Heat Tubes
Motherboard PRO Z890-S WIFI MSI Intel Z890 chipset, LGA1851 socket, Wi-Fi, Intel 200S Boost support
Memory CP2K32G60C40U5W Corsair 64 GB (32 GB ×2) DDR5, 6000 MT/s, CL40, support Intel XMP 3.0 and AMD EXPO
Storage CT2000T500SSD8JP Crucial 2 TB NVMe SSD, PCIe Gen4, high-speed M.2 storage
Graphics Card GeForce RTX 5060 Ti 16G
VENTUS 2X OC PLUS
MSI NVIDIA GeForce RTX 5060 Ti, 16 GB GDDR7, factory overclocked, dual-fan design
PC Case North Charcoal Black TG Dark Fractal Design Mid-tower case, tempered glass side panel, airflow-focused design
Power Supply AG-650M-JP *1 Apexgaming 650 W, 80 PLUS Gold certified, fully-modular PSU

System Notes / Conditions

  1. Graphics Card Power Headroom: A 650W power supply is sufficient for standard operation, but when placing high loads on both the CPU and GPU simultaneously, stability may be reduced compared to a higher-wattage PSU configuration (a 750W or higher PSU is planned for future replacement).

Installed and Required Software

1. LM Studio – Core Local LLM Runtime

Primary Role: Local inference engine and model manager for LLMs and multimodal models.

Models Tested

Key Features & Use Cases

LM Studio provides API endpoints to multiple downstream applications, enabling easy and stable integration of LLMs into workflows. It also supports MCP (Model Context Protocol)-based web search integration, allowing for advanced inference using online information even in a local environment. Since users can approve each internet search request individually, this maintains the advantage of local LLMs in controlling data leakage.

2. Kokoro-FastAPI – Local Text-to-Speech

Primary Role: High-quality local text-to-speech generation.

Stack

Features & Use Cases

This setup allowed fully local TTS without reliance on cloud APIs, with acceptable latency and consistent audio quality.

3. n8n on WSL (Dockerized Automation)

Primary Role: Workflow automation and orchestration.

Integrations

Use Cases

I found that the publicly available Docker Compose YAML for n8n can be deployed smoothly and reliably in a WSL environment with little to no modification. I also verified that—using the same node configuration—workflows can be tested by substituting external APIs with a local LLM endpoint, which proved especially useful for prototyping purposes.

4. Visual Studio Code – AI-Assisted Development

Primary Role: Local-AI-powered code editor.

Features & Integration

Use Cases

This setup demonstrated that a fully local AI coding environment is achievable for many everyday development tasks. At the same time, it became clear that 16GB of VRAM is insufficient for utilizing advanced LLM models, and that even simple coding tasks require more VRAM due to the importance of context length. Considering GPU performance and pricing as of December 2025, relying on external services for AI coding is likely the most practical approach in most cases.

5. ComfyUI – Image and Video Generation

There are three ways to install ComfyUI in your local environment: 'Desktop Application', 'Windows Portable Package', or 'Manual Installation'. Prioritising flexibility this time, I opted for 'Manual Installation' and tested several workflows. For a more casual approach, I recommend the 'Windows Portable Package', which can be launched immediately after downloading.

Environment

Models and Workflows Tested

Use Cases

ComfyUI's flexibility makes it ideal for experimentation, but careful management is required for environment control and dependency management (such as Python versions and CUDA). Additionally, GPU memory limitations become particularly noticeable during video generation.

Overall Observations