Which Qwen3.8 27B Quant Should You Run? RTX 5090 vs DGX Spark

Published 29/09/2026

Description

Which Qwen3.8-27B model should you download for local AI and coding? In this hands-on comparison, I test Qwen3.8 quantizations across an RTX 5090 desktop, an RTX 5090 laptop, and an NVIDIA DGX Spark—using LM Studio, Claude Code, and a repeatable C# coding task. Contents: 0:00 - Introduction 0:10 - Description 4:52 - Hugging Face 7:11 - LM Studio Download 11:08 - Model Tuning 15:44 - Chat Template 19:27 - Context Window 21:41 - RTX 5090 32gb Chat(s) 26:45 - RTX 5090 32gg Claude 36:19 - RTX 5090 24gb Mobile 39:51 - Dgx Spark 50:46 - Conclusion I start with the choices behind a download: GGUF versus other model formats, runtime selection, VRAM versus unified memory, and the trade-offs between IQ1, Q2, Q3, Q4, Q5, Q6, Q8, and BF16. Then I show why model-file size is only part of the calculation: context length, KV-cache quantization, batch sizes, and parallel sessions also affect whether a setup fits and how fast it runs. Unsloth publishes Qwen3.8-27B GGUF and NVFP4 variants, while its model guide lists the substantial memory difference between lower-bit quants and BF16. The video goes beyond a simple tokens-per-second comparison. I test whether each configuration produces a working .NET application, show the chat-template and context-length issues I encountered when connecting Claude Code to LM Studio, and compare the practical trade-off between faster GPU inference and the DGX Spark’s larger unified-memory capacity. In my tests, the smallest quant was not the most useful: IQ1 failed the coding-agent task, while higher quants produced working projects under different time and memory constraints. The BF16 run on the DGX Spark took much longer and still had a runtime issue in the generated project. These are results from my setup and prompt—not a universal ranking. If you are choosing a Qwen3.8 model for local coding, watch the tests before assuming the largest quant, fastest token rate, or smallest download will give you the best result. #Qwen38 #LocalLLM #ClaudeC