Qwen 3.8 Flash Next is scary good
A model that runs well on 120GB of RAM that covers 90% of my needs
September 14, 2026
Qwen 3.8 Flash Next NVFP4 has already won me over for simple devops stuff, with the Brave Search API it is essentially a family ChatGPT replacement, but today I gave it a pretty fun problem at it returned a great solution back in only 30 minutes.
I wanted Reddit Answers in my Open WebUI as a tool. I started by giving Qwen via pi.dev the reddit.com/answers URL and asked it to write a python script to ask questions and return an answer. It figured out session handling for unauthenticated usage, the quota system, and had a good working 400ish line solution within 20 minutes. Impressive!
I then asked it to integrate it into Open WebUI as a tool. No more prompting than that, no suggested solution. It weighed workflows vs pipelines, settled on workflows, realized it had to make the python async then built that. Done and tested in 10 minutes by the pi.dev agent and it worked like a charm:
Reddit Answers grounding to questions asked is a pretty nice feature. Now isn't a super complex task, it easily fit into the 256k token context window, but I am still very impressed. Pi figured the solution out, wrote code and had it running within 30 minutes. All running on a 100 watt DGX Spark*, humming at 40+ tok/s. The future is exciting.
* Even with their modest power draw the Sparks do get kind of warm, but a $30 Thermaltake "Massive 14" laptop cooler will cool a pair of them down to 66C under sustained load!