Local coding agent with Asus DX10

Accelerated Computing DGX Spark / GB10 User Forum DGX Spark / GB10
1

Hi guys.

I’ve just got an DX10 and I want to have it power the local side of my hybrid vibecoding scenario in which GitHub’s Copilot may create a plan for developing the features and the LLM model on DX10 will then take that plan and execute on it.

So far I tried the Ollama container as per Open WebUI with Ollama | DGX Spark and I was able to add it to VSCode’s Chat (Continuum) and the hybrid workflow worked but the local output was painfully slow.

I then tried sglang with help from Google’s Gemini :) but I could not find good container images and installing it directly on the hardware with Qwen2.5-Coder-32B-Instruct it was a whakamole-like experience chasing ever older version of python that would support version of sglang with the right version of torch and the right version of flashinfer… Hours and hours of installing stull only to get at best 6.something tokens/second.

Oh, and VSCode’s Chat would then not be able to communicate with sglang because it apparenyl only wants ollama as ‘vendor’ for local models and I could not make litellm to trick it into thinking it is talking to one :)

So, as phase 1, does anyone have a good working stable solution for running at peak performance a coding model like Qwen2.5-Coder-32B-Instruct?

2

If you are just getting started I recommend you begin with using https://sparkrun.dev/ to manage inference models on your Spark. For local coding model recommendations, you can review the forum but in general qwen-3.6-35b model recipes are fast and good, qwen-3.6-27b are slower and better in some cases and qwen-3.5-122b (various quants) is probably the best bet for a single spark. There are various projects which allow you to run this model at 50tok/s, which is no slouch. GitHub - albond/DGX_Spark_Qwen3.5-122B-A10B-AR-INT4: Qwen3.5-122B-A10B on DGX Spark: 28.3 → 51 tok/s (+80%) · GitHub