Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

Story

Developer details Lamplight, a local-first AI harness running two machines

AI summary

A developer detailed Lamplight, a custom harness built on Pi-Bolt plus custom extensions that links two machines — a MacBook Pro (M5 Max, 128 GB unified memory) and a self-built small-form-factor PC (RTX 5090 FE, Ryzen 7 9800X3D, 64 GB RAM) — so one can orchestrate the other via git worktrees and an event-driven job system, with Claude also involved. He says roughly 95% of his AI use now runs locally, with both machines running Qwen 3.8 Flash-Next at a reported 262K context, averaging 302 tok/s on the RTX 5090 machine (Basalt, IQ3_S) and 100 tok/s on the Mac (mlx-serve, 4.7bpw MLX), with HumanEval scores of 99.4%/93.9% and 98.8%/94.5%, LiveCodeBench v6 of 85.7% and 85.1%, and MMLU-Pro of 87.9% and 90.7%. He reports that undervolting the 5090 with a custom fan curve (575 W, 83 °C down to 471 W, 70 °C) and capping the Ryzen clock at 4.8 GHz during decode (120 W, 96 °C down to 77 W, 70 °C) fixed thermal problems, and that a "medium" reasoning setting on the Mac avoided output-cap stalls. The harness itself is not being shared yet, though the developer says he is decoupling parts of it to release components later.

In-depth insight

The post is a detailed firsthand account of running AI mostly on consumer hardware, with specific power and thermal tuning numbers and quantization choices that local-first builders can compare against their own setups. Because the harness is unreleased and all figures are self-reported by one author, the setup cannot yet be independently verified or reproduced.

#Lamplight#Qwen 3.8 Flash-Next#Pi-Bolt#Basalt#mlx-serve#Claude#RTX 5090#Apple MacBook Pro M5 Max#AMD Ryzen 7 9800X3D#Fractal Terra#MLX#llama.cpp

Original · 4 min readRead the original
Group chat

No comments yet. Start the conversation.