Claude Code + LM Studio: Qwen Coding Models on RTX vs DGX Spark

Published 28/07/2026

Description

Run Claude Code locally with LM Studio and switch between different local LLMs for AI-assisted development. Contents: 0:00 - Introduction 0:13 - Description 3:05 - Claude - Terminal 6:12 - Claude /Model 9:22 - RTX vs DGX 12:21 - LM Studio 0.4.20 Tweak 14:32 - The Prompt 15:21 - RTX Qwen 17:29 - RTX Ornith 18:24 - RTX Agents-A1 22:06 - DGX Qwen 23:36 - DGX Ornith 25:22 - DGX Agents-A1 27:22 - Stats 28:51 - Conclusion In this video, I compare several local coding-model configurations while using Claude Code as the coding-agent interface. I test Qwen-based models and coding-focused variants across two very different local AI systems: a fast RTX-based workstation with limited VRAM and an NVIDIA DGX Spark with much more unified memory but lower inference speed. The central question is simple: which local model and hardware combination gives the best practical coding experience? A fast GPU can load and generate quickly, but its available memory can limit model size and context length. A DGX Spark can handle much larger models and contexts, but generation speed may be slower. The best result depends on the task, model behaviour, tool-use reliability, context requirements, and how many retries the coding agent needs. In this video, I cover: Connecting Claude Code to LM Studio Using a local OpenAI-compatible API endpoint Configuring local API authentication Starting Claude Code with a predefined local model Creating command scripts for different LM Studio servers Passing resume-session IDs when restarting Claude Code Mapping local models to Claude Code model slots Switching local models from within Claude Code Running local models on an RTX-based workstation Running larger local models on NVIDIA DGX Spark Comparing Qwen, coding-tuned, and agent-focused model variants Testing tool use, code generation, retries, and build results Working with large local context windows Understanding why tool reliability matters more