Blog

Notes from production — inference, document AI, and the parts of a system that only show up once a model meets real hardware.

September 14, 2026
Running DeepSeek-V4-Flash across four free GPUs

A runbook for serving DeepSeek-V4-Flash on 4× L4 when the model is nearly twice your total VRAM — seven steps, with every FreeToken flag, startup log line and real error explained.
- inference
- mixture-of-experts
- gpu
- deepseek