Dev open-sources inference-optimisation agent skill he says made him $100K

Kyrannio · x · 2026-11-17

Developer @0xSero open-sourced Sybil Solutions' MIT-licensed inference-optimisation skill, claiming it has earned him at least $100K. The agent skill hill-climbs local LLM serving speed on any hardware and quant, using ceilings-first, one-change-at-a-time, interleaved controls and full-context stability gates. Reported results: 2.1× faster 256K prefill (297→628 tok/s on Arc B390 iGPU), 4× larger 256K KV cache (264K→1.05M tokens within 46 GiB), 53 tok/s single-user decode on laptop iGPU (103 tok/s with 4 users), cold start cut from 113s to 43s with 130ms TTFT. Also: Arc Pro B70 at 91.2 tok/s (3.6× vs llama.cpp 25.0), RTX 3090 up to 358 tok/s with MTP, 4×H200 at 229 tok/s under load. Repo: github.com/sybil-solutions/inference-optimisation-skill

Original post →

More from coding & agent

coding & agent channel →