Tesla Roadster Thrusters and Meta's New Open-Weight Model

Tesla showed a Roadster doing a flip using rocket thrusters. Meta released its biggest open-weight AI model and still came in behind China's GLM 5.3 Max and Kimi K3. Only one of those two changes anything for your business this quarter. It is not the car.

The Tesla Roadster rocket thrusters demo

At the Roadster event, the car first stuck itself to an upside down ramp using air thrusters. That alone would have been a fine evening for most car companies.

Then it drove at a line of Teslas and Cybertrucks. It fired what looked like SpaceX thrusters, jumped, and flipped. The same tech family that lands rocket boosters is now bolted to a sports car.

Sam Altman is not exactly on Elon Musk's Christmas card list. He said the new Roadster looked amazing. That is probably the fairest review available.

Now the boring part. The Roadster was announced back in 2017. The reported price is around $200,000. Production is reportedly still a couple of years out, around 2028.

So a car that can backflip is nearly solved. A car that arrives on the quoted delivery date remains an open research problem.

The critics called it impractical. They asked whether it will ever be street legal. They have a point. It is also one of the coolest engineering demos in years.

Both things are true at the same time. That happens more often than the internet allows.

Here is why I'm writing about a car on an AI blog. This is the shape of almost every AI demo you will see this year.

A stunning video. A huge claim. A ship date far enough away that nobody remembers the promise. The physics works, and the thing you can buy today is still nothing.

So when a vendor shows you something jaw-dropping, ask three boring questions. What can I buy right now? What does it cost? What breaks?

If the answer is "we're targeting 2028", enjoy the video and keep your money. You can admire the flip without putting down a deposit.

Meta's open-weight model vs GLM 5.3 Max and Kimi K3

On 15 October, Meta released a new open-weight model. Open weight means the files are downloadable. You can run the thing on your own hardware. No paying per call to an API.

It's huge. Around 816 billion parameters in total. About 42 billion of them are active at any one moment.

That's a mixture of experts design. Only a slice of the model fires for each piece of text. But you still need the whole thing sitting in memory.

Free to download, yes. Free to run, only if you own a rack of GPUs.

Meta is calling it the most capable US open-weight model. Read that carefully. It is a ranking with a nationality filter attached.

On the wider benchmark tables, GLM 5.3 Max and Kimi K3 are still ahead on most tests. I can also be the best AI shop on my street. I just have to pick a short enough street.

Credit where it's due. It scores very well on agent and job automation benchmarks. That's the part that matters if you want software doing multi-step work.

The license is reasonable for normal companies. Hosting providers above $20 million in yearly revenue need a separate deal with Meta. If that's not you, you're fine. If you resell AI services, read the license anyway.

The bigger story is who owns the open layer now. Qwen, from Alibaba, has been downloaded over two billion times. More than 150,000 other models are built on top of it. That's the default base layer for a lot of the world's AI tinkering.

Mistral in Europe is also going open. That's what you do when you're behind. You trade control for distribution.

And the gap to the paid frontier models is shrinking fast. On one popular intelligence index, the top open-weight model reportedly sits near Sonnet 5.5 at high effort.

Treat that as a claim rather than a law of nature. But if it's roughly right, the free-to-run stuff is months behind the expensive stuff. Not years.

When open weights are worth the hassle

Most small businesses should keep paying for a hosted API. It works. It's cheap at low volume. And nobody on your team babysits a server at 2am.

Open weights get interesting in a few cases. Your data genuinely cannot leave your building. Your volume is high enough that per-token pricing hurts. Or you want insurance against a provider changing prices, limits or terms.

If one of those is true, you don't have to build a data centre. Rent an open model from a hosting provider. Keep the option to move.

And please stop choosing models off leaderboards. The order changes every few weeks. The tests are not your business.

Take 20 real tasks you'd hand to a model. Write down what a good answer looks like. Run them through two or three options.

That takes an afternoon. It tells you more than every benchmark chart combined. If you're putting AI agents into a real workflow, that afternoon is cheap insurance.

My honest read on today is simple. The model you can download and test this week beats anything with a 2028 delivery date. One of them flips over a Cybertruck. The other one answers your emails.