# Single Spark Dev Setup **URL:** \ **Category:** DGX Spark / GB10 Projects\ **Tags:** agentic-ai\ **Created:** [October 16, 2026, 6:18pm UTC](https://forums.developer.nvidia.com/t/single-spark-dev-setup/386075 "2026-10-16T18:18:53Z")\ **Posts on this page:** 6\ **Page:** 1
**Author:** ![john608](https://developer.download.nvidia.com/images/forums/profile-default-devtalk-84.png) [@john608](https://forums.developer.nvidia.com/u/john608)\ **Post date:** [October 16, 2026, 6:18pm UTC](https://forums.developer.nvidia.com/t/single-spark-dev-setup/386075/1 "2026-10-16T18:18:53Z")
Those who use their spark for agentic development workflows what is your setup? I have tried all the inference engines and all the models and keep switching hoping to find one that stands out. I think right now Flash Next is where it is at, does anyone else have a different opinion? I have tried the Azampatti INT4 that has the reduced experts and that has given me the most consistent speed, Mia Tensorfold is pretty cool too but had some tool calling issues. I’m currently testing Athena Inference Engine which is awesome for wanting to fast switch between flash next and ds4 flash but I find the speeds aren’t what I was hoping for. ---
**Author:** ![door.blu](https://developer.download.nvidia.com/images/forums/profile-default-devtalk-84.png) [@door.blu](https://forums.developer.nvidia.com/u/door.blu)\ **Post date:** [October 16, 2026, 10:18pm UTC](https://forums.developer.nvidia.com/t/single-spark-dev-setup/386075/2 "2026-10-16T22:18:00Z")
Try just released nvidia flash-next-fp4. ---
**Author:** ![john608](https://developer.download.nvidia.com/images/forums/profile-default-devtalk-84.png) [@john608](https://forums.developer.nvidia.com/u/john608)\ **Post date:** [October 16, 2026, 10:53pm UTC](https://forums.developer.nvidia.com/t/single-spark-dev-setup/386075/3 "2026-10-16T22:53:06Z")
I’ll give it a look, what is the main advantage to it over nvfp4? ---
**Author:** ![luisiam](https://developer.download.nvidia.com/images/forums/profile-default-devtalk-84.png) [@luisiam](https://forums.developer.nvidia.com/u/luisiam)\ **Post date:** [October 16, 2026, 11:50pm UTC](https://forums.developer.nvidia.com/t/single-spark-dev-setup/386075/4 "2026-10-16T23:50:28Z")
I think Qwen3.8 Flash Next is the best model for a single spark as long as PLE can be offloaded to the NVMe. Qwen3.8 27B will be quite a bit slower. Take a look at the flash next FP4 checkpoint recently released from NVIDIA. Based on my understanding it is still NVFP4 but a bit more aggressively quantized such that it run decently on a single spark. QAD is applied to compensate the quality loss in quantization. ---
**Author:** ![john608](https://developer.download.nvidia.com/images/forums/profile-default-devtalk-84.png) [@john608](https://forums.developer.nvidia.com/u/john608)\ **Post date:** [October 18, 2026, 2:32am UTC](https://forums.developer.nvidia.com/t/single-spark-dev-setup/386075/5 "2026-10-18T02:32:48Z")
Yes I’m currently comparing the FP4 release to the NVFP4 with PLE offloaded and also trying to compare it to @azampatti INT4 version that has less active experts. ---
**Author:** ![huggface](https://developer.download.nvidia.com/images/forums/profile-default-devtalk-84.png) [@huggface](https://forums.developer.nvidia.com/u/huggface)\ **Post date:** [October 26, 2026, 6:01am UTC](https://forums.developer.nvidia.com/t/single-spark-dev-setup/386075/6 "2026-10-26T06:01:59Z")
Do yourself a favor and just run this Cockpit LLM manager. Once installed, it gives a web dashboard to install the most optimized versions of LLM. I’m running qwen-3.8 flash - nvidia release. Running it for days on coding large project. Averaging 32 tok/s. Local llm now has the ease of frontier , with token speed that make it worth running local. ![image](https://global.discourse-cdn.com/nvidia/original/4X/4/b/5/4b5f2341678ef99dec10e230a11ebc73be9084f8.png) > **[GitHub - hasso5703/dgx-spark-qwen38: One command turns a DGX Spark (64 or 128 GB) or an...](https://github.com/hasso5703/dgx-spark-qwen38)** > > One command turns a DGX Spark (64 or 128 GB) or an ASUS Ascent GX10 into a private AI server: nine measured, pinned lanes (Qwen3.8 27B and Flash up to 1M tokens, Ornith, Nemotron, Kolibri, images, video with sound), OpenAI and Anthropic APIs, and a local cockpit on a laptop or a phone. Nothing leaves the box.