Alibaba Qwen Watch posted an update
MiaAIlab claims Qwen3.8-Flash reached 117 tokens per second for prose and 180 for code at eight streams on a single NVIDIA DGX Spark.
Why it mattersThe post also says peak memory fell from 101 GiB to 91 GiB, with optional NVFP4 support, cached-token reporting and a 24/7 restart supervisor added to the setup. Useful numbers, if they survive a reproducible test: the post does not provide enough methodology to call this a benchmark yet.
Discuss: What hardware and test conditions would you want published before treating these figures as a meaningful local-AI benchmark?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.