Lane 01
Local inference, private agent systems, runtime architecture, and the hardware constraints that refuse to stay inside neat diagrams.
Nineteen days after publishing Pennyroyal, another user reported 4.8 billion prompt tokens, 55 million generated tokens, and more than a week of real work on a 300 W Max-Q. Then a second report arrived from the 27B configuration.
Read the field report
Several speculative paths moved Qwen3.8 from 74.06 to 108.75 tokens per second on one RTX PRO 6000. A 29 percent long-prefill regression then sent the runtime through HiCache and NIXL so it could stop rebuilding prefixes it already knew.
Read the runtime field note
My RTX PRO 6000 can consume 600 watts, but local AI rarely asks for all of it. A tuned 450-watt profile preserved practical Qwen performance, while a completed 994,987-token DeepSeek V4 Flash request averaged just 202.78 watts.
Read the hardware field note
A 96 GB RTX PRO 6000 led to transient-power reboots, a third power supply, dual Xeons, external GPUs, a Dremel, blood, and a 1,500-watt MiniMax M3 experiment that still came up six gigabytes short.
Read the hardware field note
The Radeon AI PRO R9700 installed cleanly, ROCm worked immediately, and Vulkan proved 74% faster at prompt ingestion and 32% faster at token generation in the cleanest early comparison. The harder failures came from the stack around it.
Read the field note
DeepSeek V4 Flash fit until its speculative decoding head pushed the system out of memory. Five CUDA-mapped expert layers recovered enough VRAM for DSpark-4, the FP4 correction tier, and a practical near-million-token context.
Read the technical field note
The origin story, the hardware and runtime escalation, the correction loop, and why the model is only one part of a useful private AI system.
Read the field note