Gary Vaynerchuk once said he gets more joy from finding and flipping hidden gems at garage sales than from closing million-dollar business deals.
This seems irrational from an ROI or impact perspective (after all, who cares if you found an original McDonald’s Hamburgerler Happy Meal Toy?!). However, it’s hard to put a price on things that make us happy for no reason.
I had two such experiences last month testing the frontier of local AI.
Experience 1: “Wait, so this is running at your house?”
Besides being an amazing human being, Bryan is one of the most mission-driven founders you’ll ever meet. He’s been building the HAX CMS tools and community for over a decade, and it’s been amazing watching his journey.
AI would be a perfect tool to uplevel his capabilities, but it’s hard to pay for Fable on a professor’s salary and university budget. So, like many, he tries to stretch his personal funds without breaking the bank.
Most of us AI-pilled folks are kinda spoiled in this regard. We don’t hesitate to have multiple $200/month subscriptions to test and use ALL THE THINGS!!!! However, for many people, this kind of expense is a family budget discussion (think single moms, college students, folks in India, etc.).
So I decided to try and experiment. What if I opened up my extra inference (specifically with GLM 5.3 Flash) on my two DGX Sparks to Bryan? It wasn’t as simple as I wanted, but Antseed recently released a tool to do just that.
I waited with Christmas-morning-like anticipation to see Bryan’s reaction, and was thrilled to get this DM.
“Wait, so this is running at your house?”
Sure, to you this may seem as boring as Gary V hitting the garage sale. To me, it’s everything.
These are the types of mini-wow moments I obsess over as a serial startup founder. It signaled something profound to me: the ability to serve strong-enough inference in a local, shared, distributed fashion. As a heavy Napster and BitTorrent user in college, I was hooked.
This was a life-changing moment for me more than for Brian. I wanted more.
Experience 2: The New Token Maxxing: Serving Vs Consuming
Within an hour of Bryan’s DM, I knew I had to open up the pipe to see if I could saturate demand. Using Antseed’s decentralized market, I could easily flip that switch on by asking Claude to set up a DigitalOcean droplet to authorize and route traffic securely.
I flipped on the switch and... nothing. Hmm. Ok.
Then I went to bed, woke up, and refreshed.
After 8 hours, I had served over 2 million tokens to 8 people! This is more tokens than I use on a slow day, so I was encouraged!
After 2 more hours, I had my “oh shit” moment. Now I was serving 3.2 million tokens PER hour. After some napkin math, I realized that was 2 billion tokens per month!
I remember Kieran (Every team member and Compound Engineering creator) bragging about using a billion tokens in a month. To me, that felt like an impossibly high number I would never reach.
But on the SERVE side, I could passively contribute TWICE that to the Antseed network’s overall inference demand. In short, I made an impact while actively working AND passively letting my infrastructure serve inference.
How Did I Even Get Here?
A year ago, I had zero interest in running local models. I didn’t think it was economically feasible or technically possible.
But a long chain of events brought me here.
November 2025: Opus 4.5 made longstanding dev tasks possible.
December 2025: A colleague and I decided to build a startup using screen recordings
February 2026: Early user pushback was that local AI was a requirement, not a nice-to-have.
April 2026: Attended an Ollama meetup in SV where Gemma 4 was delivering Sonnet 4.5 speed and performance. On the flight home, I used Gemma 4 on my laptop with no wifi and achieved some wins.
July 2026: I start seeing the ODS team and 0xSero, and Mia Labs X accounts go all-in.
August 2026: Qwen 3.8 releases, and it’s the first strong coding model that works on a single Spark, so I buy one.
September 2026: I see all the top Nvidia GPUs get sold out and prices skyrocket, and I see the Spark inventory fly off the shelf. I buy them the last day they were still available for $5K (now they are $7K)
This started as a user need (personally and to power our startup app), but systematically led to more experiments, testing, and exploration.
Now I’m loco for local (sorry, #dadjoke). I still use frontier models for 60-70% of my workload, but I see a future where over 50% of my usage is on my Sparks.
Finally, Opus-Level Models
What I love about Every is that their vibe-checks include a “reach test”, which is the ultimate verdict. It doesn’t matter what the benchmarks say if your experience sucks, and you keep going back to another model.
For me, the reach test was whether I could use local models at an Opus 5.0-like level. The “like” is doing a lot of heavy lifting because obviously you can’t fan out 10 subagents or turn on /ultrathink like modes on finite hardware. But for work that MUST be local (client requirement or highly sensitive personal information), local was sufficient.
To me, this answer is an emphatic YES.
This threshold was crossed with the following cohort of models.
Deepseek 4.1
GLM 5.3
Qwen 3.8
Kimi 3.1 (although I’ll admit I don’t use this one)
GLM 5.3 is my go-to, and it’s a joy to work with. It thinks critically and over long time horizons. And when I hit its coding limits, I switch to Qwen 3.8 as a slower but more capable coding model.
Local vs Frontier? Or Local vs Nothing
I get the objection many have... bleeding-edge model or bust! But in my experience, burning through my limited number of Fable tokens on mid-level tasks sucks. I have to wait hours or days for my usage to reset. What then? That’s when open-weight models are the only game in town.
Not everyone has the luxury of token budgets comparable to their monthly salary (I just saw a few people bragging about burning through over $1M of Fable 5.2 credits in a week). In fact, Accenture just made the bone-headed move of limiting everyone to $29/month! That’s like 2 Opus 5 calls a week. What then?!
And even when budgets are healthy and robust, what if the data can’t be sent to Fable? Remember, Anthropic had 30-day retention policies even for clients with alleged zero data retention (ZDR) policies. That’s unacceptable for many.
For me, local is strongest when the alternative is zero AI. And that’s why I’m THRILLED to have permanent, infinite access to as much GLM 5.3 capacity as I want (even if it might be slower than what I get on OpenRouter).
What Next?
My journey is just getting started. Having experienced David Skok’s “mini-wow” and “wow” PMF thresholds, I’m hooked. But there’s a big difference between aspiring to change and actually doing it.
My goal is to route 50% of my token usage through my Sparks, starting by replacing my Grok Bot usage with a Hermes Desktop + DGX Sparks + GLM 5.3 combination. And with OpenCode’s v2.3 harness (getting closer to Claude Code, but not there yet), I will try to move as much of my personal work to GLM by the end of the month.
If you’re looking to get started on the local AI journey, do it! And you don’t need 2 Sparks. Smaller decision-making models (Jev alternatives) can fit on a Mac with 32GB of RAM. There are also other EA-optimized models like Bonsai that you can start with today (see the Underdog project).
PS. I used Venice AI to generate a commercial ad for local AI. Enjoy




