models on my hardware. one memory every tool shares. nothing leaves the house.
a self-hosted ai fleet on an intel arc b60, run from a desk in augusta. real benchmark numbers, a memory brain past 130,000 entries, and agents with guardrails that hold.
what it actually runs on.
no rack, no cloud. a mini pc with a workstation gpu hanging off an oculink dock, a standby box that watches it, and a shelf of storage and power. scroll to walk the desk.
Q4_K_M isn't Q4_K on Intel Arc.
generation speed for a 27B dense model across quant variants, and the 30B MoE for scale. measured on my rig with a 5120×1440 desktop sharing the gpu unless marked idle.
27B rows are Qwen3.8-27B on b10588 unless noted. 30B rows are Qwen3-Coder-30B-A3B IQ4_XS; 38.7 is the idle-card figure (42.0 with MMVQ off). Mesa 26.0.8 to 26.1.7 made no difference. raw jsonl in the arc-llm-bench repo.
the finding
the 67 q6_K tensors inside a stock Q4_K_M run matrix-vector at about 125 GFLOPS on Arc, against about 1,200 for q4_K. that is 88 ms of every 133 ms token. GDN accounts for 0.6 ms, so it is not the cause.
what to do with it
on this card, pick the quant for the compute path, not the file size. a pure Q4_K beats a mixed Q4_K_M every time, and a 30B MoE with 3B active parameters beats both.
03 the systems: Neuralis, Sentinel Prime, AURORA-X
Neuralis
the shared memory every tool and agent reads from and writes to. hybrid semantic and graph recall, self-correcting, past 130,000 memories. supersede, never delete.
open the recall graph →Sentinel Prime
autonomous desktop and fleet agents. hand it a goal; it plans, executes and reports back, with money and outbound traffic clamped by default. nuke-standby watches the whole time.
watch a run →AURORA-X
one screen over the fleet: the brain graph, thirteen tailscale nodes, cost, voice, vision and a chat that dispatches work. this is the real dashboard; the public copy runs on simulated data so you can poke at it.
open aurora-x →
every run makes the next one smarter.
the pieces worth sharing.
code, config and the real benchmark numbers. use them, fork them, check my work.
i fix other people's networks all day at an MSP. i wanted my own stack that i own end to end — models on my hardware, one memory every tool shares, and a system that heals itself before i notice it broke. nothing leaves the house unless i say so.
field notes.
agents, self-hosted infra, security review.
mail hello@dirtysouthalpha.com