Day 35 of testing DGX Spark with DeepSeek Harness & Qwen 3.8.
Encountered setup issues, power loss, and slow token speeds (10-12 tokens/sec vs. expected 60). Memory usage is high (90GB). Long processing times (88 mins) suggest potential issues. Will revisit tomorrow.
Alibaba just turned AI from a chatbot into a workforce.
And the biggest upgrade isn’t the 2.4 trillion parameters. It’s what happens after you give it a job. What Qwenwork can actually do: → Research a topic, write the document, generate images, build the webpage, create
Qwen 3.8 27B at 56tps; on 9 year old GPU btw
Nvidia V100 32GB ~$650 on EBay right now! Using Dflash 2; disabling the ECC adds some more speed too! Thinking and prose is a bit slower, but 56-63 tps in code gen! MTP runs faster for prose vs DFlash2 but slower sustained code
Lowk impressed i gave a local Qwen 3.8 my nix repo
Asked it to simplify whatever it can this was the result after 2 hours
You have reached the end of the archive
All of qwen38