Mac Studiohost
M4 Max · 64 GB
- Role
- Model server (oMLX)
- Runs
- Qwen3.8-27B · 4-bit & 8-bit
Rig — local AI
Local-first AI · at home
A small AI rig on my desk.
Three Macs on one private Tailscale network. The Studio serves open models through oMLX, so private work never has to leave home — and the other machines can use it from anywhere.
Mac mini
M2 · 8 GB
- Role
- Worker
- Runs
- Agents and jobs
MacBook Air
M4 · 16 GB
- Role
- Client, on the move
- Runs
- Chat · coding sessions
Local models served · 178M tokens · 7,337 requests
- Qwen3.8-27B-4bit3,698 req82M
- Qwen3.8-27B-8bit1,884 req64M
- Qwen3.5-9B-4bit1,147 req24M
- Qwen3-Coder-30B-A3B-Instruct-4bit371 req5M
- gemma-4-26b-a4b-it-4bit69 req1M
- Qwen3.6-35B-A3B-4bit21 req692K
Three Macs on one private Tailscale network. The Studio serves the models through oMLX; the mini and the Air reach it from anywhere and share the work.
Burn — tokens
Tokens through my AI toolsAs of 10 Oct 2026
31.1Btokens
97% re-read from prompt cache · 536M fresh in · 74M out
- Re-read from cache30.3B · 97%
- Written to cache228M · 0.7%
- Fresh input536M · 1.7%
- Output74M · 0.2%
By tool
Claude CodeAnthropic · agentic coding58% of all · since May 202618.1B
CodexOpenAI · agentic coding31% of all · since May 20269.72B
Z.ai GLMZCode app + opencode · GLM-5.39.4% of all · since Jun 20262.94B
Local (oMLX)On my Macs · Qwen3.8 and friends0.6% of all · lifetime178M
opencode · otherDeepSeek, GPT via opencode0.6% of all · since Jun 2026173M
Burning about 5B a week lately — 9.32B in the last 7 days. Cloud agents carry 99% of it; the local rig, 0.6%.
Week of 21 JunLast 7 days
The cloud agents — Claude Code, Codex and GLM — do the heavy lifting and re-read long contexts from cache all day. The local rig handles the private work. They draft; I review and decide what ships. Counted on my Mac from each tool's own usage logs; only totals leave the machine. Claude Code keeps its local logs for a limited time, so its count covers the retained history only.