Writing
Experiments I run in my own time, written up with the numbers and the code to rerun them.
- Running a 125B model on one DGX Spark, from 33 to 52 tok/sHow I ran Qwen3.8-Flash-Next (125B total, 6B active) on one NVIDIA DGX Spark, and the changes that took decode from 33 to 52 tokens a second.
- Qwen3.8-27B at 387 tokens a second on one DGX SparkHow I served Qwen3.8-27B on one DGX Spark with SGLang, NVFP4 and DFlash2: about 70 tok/s for one stream, 387 tok/s across 32 streams.
- EvenMic: making Even Realities G2 glasses a Mac microphoneHow I made Even Realities G2 glasses a system-wide Mac microphone, cut full-screen images from 9.4 s to 449 ms, and what it taught me about AI on devices.