DGX Spark local config: Qwen3-Coder-Next FP8 + Qwen2.5-Coder-3B autocomplete

[ztolley]

I wanted to share a tested local Continue setup for a DGX Spark private coding-assistant workflow:

https://github.com/ztolley/dgx-spark-qwen3-coder-next-compose

This is a community-tested reference setup, not an official vendor certification.

Model split:

• Qwen/Qwen3-Coder-Next-FP8 for chat, edit, and apply
• Qwen/Qwen2.5-Coder-3B for autocomplete

Why this may be useful:

• it is a real tested setup on DGX Spark
• both models are exposed as local OpenAI-compatible endpoints
• the main default context is 32768, with 40960 also validated as feasible
• the repo includes a Continue config example and measured memory/performance notes

Measured notes from the default stack:

• main model GPU memory: about 88.8 GiB
• autocomplete GPU memory: about 11.0 GiB
• repeated prompt with prefix caching: 2.37s
• autocomplete short completion: 1.56s

I also tested Qwen/Qwen2.5-Coder-7B for autocomplete under vLLM. It fit, but it was slower and not clearly better enough to replace 3B as the default.

If anyone else is running large local coding models with a separate autocomplete endpoint in Continue, I’d be interested in comparing setups.