My model said no. That isn't the control.
Three always-on agents behind AgentMinder. My model refused every planted email and brief; when I hijacked the agents, AgentMinder still said no.
Running a model privately is a systems decision. The weights must fit somewhere, tokens must arrive quickly enough to be useful, and the cost has to include idle hardware as well as peak throughput. These posts record what I ran and what the measurements did—and did not—support.
Begin with the local-model experiment, then follow the memory-bandwidth limit into the larger capacity question. The rental and breakeven articles compare ways to get compute without pretending that theoretical throughput is sustained utilization. Prices and versions belong to the date of each experiment.
Start here
- Meta's Muse Glimmer runs on my laptop with the Wi-Fi off
- I tried to make my laptop AI twice as fast. I found the wall instead.
- A 2.8-trillion-parameter model fits in 8 GB of RAM. It writes one token every 32 seconds.
- The GPU I need does not live in my house. So I rent someone else's and pretend it does.
- Self-hosting inference beats the API at 21 billion tokens a month. That is 8,000 tokens per second, forever.
- I wanted private AI. What I had to decide was Kubernetes.
Field notes on this topic
I ran GLM against DeepSeek on two Sparks. The headline number was the least useful one.
I reproduced DeepSeek v4.1 Flash at 103 tokens per second on my two DGX Sparks, then benchmarked it against GLM-5.3-Flash on real work.
Securing my agents starts with where they run
I already run always-on agents. My microVM experiment starts a series on controlling their environment and testing where each security protection applies.
How I moved the Field Signal newsroom into VKS on the Dell.
The Dell VKS build behind my Field Signal newsroom: Avi, storage, private Spark inference, migration failures, and verified publication.
I rebuilt Field Signal around my local AI.
I rebuilt Field Signal around a local model API on my Spark, with workers on Dell VKS and Paperclip for review. More room to build and learn.
I connected VCF Operations to my Spark. I am not keeping the setup.
I connected VCF Intelligent Assist to private Qwen on DGX Spark. The inventory query worked, but the required Log Management footprint did not fit my lab.
I bought a Spark. Now I want to put local agents to work.
My DGX Spark now runs Field Signal’s text agents. Next: practical agent workflows for VCF Operations, vDefend security reviews, and Avi application health.
My lab has an IT support team now. Nobody on it is allowed to touch anything.
A read-only AI IT support team for a private cloud lab: VCF Operations evidence, typed Jev judgments, and six bounded read-only specialists that end every investigation by asking a human. What is live, what is only planned, and the night the cluster it runs on ran out of memory.
My homelab was beefy. It still wasn't enough.
A homelab update: why I outgrew a rack of old enterprise hardware, the MINISFORUM MS-A2 I bought to replace it, what NVMe memory tiering does for a small box, and where this blog is headed next: on-prem AI workloads, agents, and how to run them safely.
VMware’s private cloud has more to do. Here’s what changed.
VMware Explore 2026 and VCF 9.1.1, explained: private AI, agent controls, platform operations, and what is available now.
The GPU I need does not live in my house. So I rent someone else's and pretend it does.
My homelab is great at everything except the one thing AI needs: a real GPU. So I rent one by the hour from a neo-cloud, pull it into the tailnet, and tear it down in a finally block. The buy-versus-rent math, priced honestly.
Self-hosting inference beats the API at 21 billion tokens a month. That is 8,000 tokens per second, forever.
Working out when self-hosting an open model actually beats a managed inference API, using published list prices. The breakeven is not a price per token, it is a sustained throughput you have to hit every second of every day.
I wanted private AI. What I had to decide was Kubernetes.
The GPU is the part everyone budgets for and the part I never needed. An open model runs privately on CPUs you already own — so the real decision is the platform underneath it.
I tried to make my laptop AI twice as fast. I found the wall instead.
A build log on speculative decoding, memory bandwidth, and why the honest answer to 'can we make it faster?' was 'not today' — plus exactly why, and why that's still a good day's work.
Meta's Muse Glimmer runs on my laptop with the Wi-Fi off
I put Meta's new open-weight Muse Glimmer 30B on my MacBook, pulled the Wi-Fi, and it kept working. A build log on local AI — and why running it yourself is an architecture decision, not a hobby.
A 2.8-trillion-parameter model fits in 8 GB of RAM. It writes one token every 32 seconds.
Two new engines run frontier Mixture-of-Experts models on consumer hardware by paging experts off SSD instead of holding them in memory. Why that works, why it is slow, why C is not the reason, what a GPU does and does not fix, and what it means for air-gapped estates and hardware refresh planning.
There is no cloud. It's a Dell workstation in a spare room.
A mid-2026 state-of-the-homelab: one Dell workstation running Proxmox, a Raspberry Pi, a bare-metal ESXi box, and twelve Tailscale devices running real websites and a fleet of agents, with no public IP and one deliberate single point of failure.
I didn't give my agents a vector database. I gave them 126 text files.
A working design for agent memory that survives the restart: one fact per Markdown file, an index loaded every session, wiki-links instead of a graph database, and the ways it quietly rots. Measured from a real store of 126 memories.
I built a machine that wrote 278 articles. Nobody read them.
Apollo published 278 articles in 14 weeks to zero subscribers. A postmortem of a pipeline that worked as software and failed as a publication.