Ditching llm-scaler: Qwen3.8-27B on my B70
Leaving llm-scaler behind, patching vLLM with GPT-5.6 Sol, and getting MTP4 working. My setup as of August 20.
Achu Mukundan
I'm a software engineer in Toronto. I build developer tools, local AI software, and audio tools.
Writing
All notesSelected projects
All projectsA Rust terminal coding agent for cloud and local models, with disk-backed sessions and subprocess extensions.
A deterministic DSP tool that stretches, granulates, degrades, loops, and exports audio in the browser.
A native macOS player with real-time spectrograms, multiband analysis, and a phase-vocoder pitch engine.