Quick game. Three ways to get somewhere by car:
- You call a cab. You get in, say where you're going, and a stranger drives. You never see the engine. You have no idea if they're taking the scenic route to run up the meter.
- You lease a car. You get the keys and drive it yourself, wherever you want. But you didn't build the engine, the manufacturer still knows exactly where that car goes every day, and if they decide to stop making the model, good luck finding parts.
- You build the car yourself, in your own garage, from a full parts list and instructions anyone can read. It's slower to get moving, but nobody else has a key, nobody's tracking the mileage, and if something breaks, you actually know which bolt to check.
That's not just a car analogy. It's the entire AI industry, and most people don't realize which one they got into this morning.
Cab, lease, or garage: pick your AI
- Closed, API-only models (the GPTs and Claudes of the world, accessed only through someone else's servers) are the cab. Convenient, and someone else worries about maintenance. But every destination you whisper to the driver, every prompt, every internal doc, every half-baked idea, goes through their dashboard, not yours.
- "Open-weight" models are the lease. You get to download the actual weights and run them on your own machines. That's real progress: no data has to leave your building to get an answer. But you're still driving something someone else designed, under whatever license terms they wrote, and you can't always see what went into building it.
- Fully open-source AI (weights, training code, and ideally the data recipe, all public) is the garage build. This is the version where "open source" actually means what it always meant for Linux or Python: not just "free to use" but inspectable, modifiable, and yours.
For years, picking option 2 or 3 meant accepting a worse car. The cab company simply had better engines. That gap just got a lot more interesting.
The garage just rolled out a supercar
On July 16, 2026, Chinese lab Moonshot AI unveiled Kimi K3, a 2.8-trillion-parameter model it's calling the largest open-weight AI system ever built, with full weights following shortly after.
Here's the fun part, and it's very on-theme for a car analogy: Kimi K3 doesn't run all 2.8 trillion parameters at once. It's built as a sparse mixture-of-experts, with 896 "expert" sub-networks under the hood, but only 16 of them fire for any given token. Imagine an engine with 896 cylinders that only fires 16 of them per stroke, and somehow that's the fuel-efficient part. The catch: Moonshot itself recommends deploying it on "supernodes" of 64 or more accelerators, so this particular garage build is really more of a datacenter build. It's not landing on your laptop anytime soon.
The garage build isn't just cheap to fuel; it's fast too. On the independent Artificial Analysis Intelligence Index, K3 landed as roughly the third-strongest model family overall, close behind GPT-5.6 Sol Max and Claude Fable 5, and ahead of Claude Opus 4.8. On Arena's Frontend Code leaderboard, judged blind by real developers, it actually took first place, ahead of Claude Fable 5. A car you can build yourself, in your own garage, just out-cornered one of the fanciest taxis on the road.
Reading the fine print before you celebrate
No test drive is complete without kicking the tires:
- Some of the flashiest claims are Moonshot grading its own homework. The core leaderboard rankings above come from independent evaluators, which is reassuring, but splashier claims (like a reported 48-hour autonomous run where K3 designed a simulated inference chip) come straight from Moonshot's own blog and haven't been independently reproduced. Healthy skepticism applies to open models exactly as much as closed ones.
- "Open" doesn't mean "runs on a laptop." Even with sparse activation keeping compute lower per token, all 2.8 trillion parameters still have to be stored somewhere, which is why Moonshot points people toward serious multi-accelerator hardware. You're building the car yourself, but you still need a real garage, not a driveway.
- Provenance is a live argument, not a settled one. Anthropic has publicly alleged that Moonshot trained on a large volume of Claude conversations through distillation, a claim Moonshot disputes. It's worth tracking, not worth treating as resolved either way.
- A great benchmark score doesn't hand you a maintenance contract. Winning a leaderboard doesn't automatically solve the procurement, support, and compliance questions that come with betting real infrastructure on a new vendor.
The big takeaway
For a long time, "open source AI" was a purity argument: nice in principle, worse in practice. What's changing isn't the principle. It's the practice.
The gap between "I built this in my own garage" and "I'm renting the fanciest cab in town" is shrinking fast enough that it's now a genuine engineering trade-off, not a consolation prize. You give up a bit of convenience, and in return you keep every mile you drive to yourself.
Just remember, buying the parts list doesn't make you a mechanic overnight. Building the car is only step one; you still have to keep it maintained, insured, and out of the neighbor's fence. But for the first time in a while, it's a car actually worth building.
Sources