Of course, there may be people who can use it better than me,
but for now, I've settled on the DeepSeek V4 flash 0731 model.
Using vllm and Dspark, the speed is about 75~80 tok/s on average.
Since there are many problems that require deep digging rather than simple question answering, I've set thinking to max.
I've also linked a Telegram bot, so it seems sufficient as a personal service.
The button in the middle allows you to switch between LLM mode (vllm, open-webui) and video mode (comfyui minimax-h3).
Actually, I made it so that status monitoring and service switching are possible through webpage buttons because it was a hassle to connect remotely.
The screenshot is of the service page. (Although, this was also created by an llm and I just followed along)
For video mode, I'm considering setting up the qwen 3.8 27b model since minimax h3 doesn't require two dgx sparks. (I'm thinking about fine-tuning the qwen 3.8 27b model...)
It is certain that the dgx spark is a very cost-effective device, but looking at the recently released open weight models,
it's not without its subtleties.
Still, I think the deepseek v4 flash 0731 model could be a custom killer model for two dgx sparks.
The biggest advantage is that it allows you to try processing commercial data locally (although deepseek can't solve it on its own). I believe that better open weight models will come out over time, so I'll leave a simple usage guide.
