Best models/configurations for agentic coding with DGX (Nvidia/Asus/Dell/Lenovo/MSI)

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

It has been about a month that I use a PGX (Lenovo version) for agentic coding with OpenCode mainly (also trying Goose).

I tried multiple models with different engines (vLLM, SGlang) and different tools (‘made’ by members of this community). On this I really want to thank many of you for your work! Without you, your dedication, your time, using the DGX would not be easy (this product is great, but not ready for mass deployment. At this time, it feels more as an AI/IT lab tool).

It’s difficult to test everything in ‘work’ conditions. Benchmarks are great but I’m more interested in quality that rapidity so 15 t/s for ‘Claude’ like quality could be fine for me.
Also trying the same model with vLLM (eugr qwen3-coder-next-fp8) or Sglang (sparkrun sglang qwen3-coder-next-fp8) with OpenCode, it feels ‘different’ on the reasoning.

What is best for short code / long refactoring ?

I open this subject for sharing/discussing our experiences…

2

For me,

At this point the best ‘experience’ quality/speed is using

Model: Qwen3-Coder-Next-FP8

Engine: SGlang

Running with sparkrun is the easiest:
sparkrun run qwen3-coder-next-fp8-sglang --tp 1 --solo

OpenCode config:

…
“agent”: {
“build”: {
“temperature”: 0.7,
“top_p”: 0.95
},
“plan”: {
“temperature”: 1.0,
“top_p”: 0.95,
“repetition_penalty”: 1.12,
“presence_penalty”: 0.6,
“frequency_penalty”: 0.5
}
}
…

3

Here is a great start :
Bfloat16 Quality = Speed? - DGX Spark / GB10 User Forum / DGX Spark / GB10 - NVIDIA Developer Forums