Meta's Two-Track Bet, and a Roadster That Hops

Two features this week, because two stories earned it.

The first is Meta. On Monday it released Muse Wonder, its most capable model, closed and sold through an API. On Thursday it released Muse Spark Open, an 816-billion-parameter model whose weights anyone can download. Same company, same week, opposite choices. I wanted to understand what's actually different, and what happened to the Meta that used to argue open source was the future.

The second is a car that jumped over five other cars on Thursday. I have questions.

Meta's two-track bet

How we got here

The short history, as I've pieced it together:

• Llama (2023–2025). Meta's reputation in AI was built on downloadable weights. Llama 2 allowed commercial use. Developers built a whole ecosystem on it.
• Muse Spark (Apr 8). The first model from Meta's rebuilt Superintelligence Labs, and the break. The Register's summary was "this is no Llama": no weights, access through Meta's app or an invite-only API.
• Muse Code and Glimmer (August). Muse Code, a terminal coding agent, launched Aug 5 with a "Contributor" tier that, per MacRumors, costs up to 20 times less if you let Meta train on your prompts and completions. On Aug 10 came Muse Glimmer, a 30B open-weight model under Apache 2.0, small enough for a single consumer GPU. Reports say it was distilled from the closed Spark line.
• Muse Wonder (Oct 12). Meta's "most capable models to date," aimed at agentic work in coding, law, finance and health. Available through the Meta Model API and Muse Code. VentureBeat lists the API price at $4 per million input tokens and $20 per million output. No weights.
• Muse Spark Open (Oct 15). Meta calls it "the open-weight version of Muse Spark" and "the most capable US open-weight model to date."

Put those in a line and the pattern is clear: the frontier stays closed, and the tier below gets given away.

What's technically different about Spark Open

From Meta's post and the model card coverage:

• 816B total parameters, 42B active. It's a mixture-of-experts (MoE) model. Instead of every parameter working on every token, a router sends each token to a small subset of "expert" sub-networks. You get the knowledge capacity of a huge model with the per-token compute of a much smaller one.
• The catch with MoE: active parameters cut compute, not memory. All 816B parameters still have to be loaded somewhere. VentureBeat cites the model card at at least 700 GB of accelerator memory for the NVFP4 build and 2 TB for BF16. Meta's own demo ran it across a pair of top-end desktop machines. For me, on a 32 GB M1 Max, this is a hosted-API model, not a local one. Glimmer is the one built for laptops.
• 1M-token context, and Meta describes it as natively multimodal: text, image, video and audio in, text out.
• Extra safety post-training before release, because once weights are out, there's no server-side guardrail to fall back on. Meta says it met the thresholds of its own Superintelligence Scaling Framework, with a full evaluation report "available soon."

What we don't know is how much Spark Open differs from the hosted Spark models. Meta calls it the open-weight version, but hasn't said whether it's identical.

Open weights, not open source

This is the part worth slowing down for. "Open" is doing a lot of work in that name.

The weights ship under a custom Muse Model License, which Meta itself describes as having "limited commercial terms." From VentureBeat's reading of it:

• You can run, modify, fine-tune, distil and redistribute it, subject to Meta's usage policy.
• Products must show "Built with Muse Spark Open."
• "Derivatives" are defined broadly, including models improved using Spark Open's outputs.
• If you sell model access as a service and pass $20 million in revenue over 12 months, you need a separate agreement with Meta first.
• Downloading means filling in a Hugging Face form asking for your legal name, date of birth and organization.

None of that is evil, and for most app developers it won't bite. But it isn't open source in the sense the Open Source Initiative means it: there are use restrictions, a revenue trigger and an attribution requirement, and you get weights, not training data or the full recipe. "Open weights" is the honest term.

Compare that with Glimmer, under plain Apache 2.0. Meta now ships three tiers: the small model with a real open-source license, the big model with strings, and the frontier model with no weights at all.

How I read the strategy

The way I understand it, each tier does a different job:

• Glimmer wins developer goodwill and the local-agent niche.
• Spark Open keeps Meta in the "best US open model" conversation, and the license makes sure the biggest commercial reusers come and talk to Meta.
• Wonder is the product. It earns API revenue, and Muse Code's cheaper tier earns training data.

That's not unusual anymore, but it's a long way from Meta's old argument that openness was the strategy. Openness is now a tier.

The Tesla Roadster that "flies"

On Thursday, at SpaceX's McGregor, Texas test site, Tesla finally showed the second-generation Roadster, nine years after it was first announced. Here's what Tesla and the press reported:

• A prototype drove up a huge inverted, curved ramp, held on by a suction system, then used cold-gas thrusters to leave the ramp, spin, and land.
• It also lifted itself over a Cybercab, Cybertruck, Semi, Model Y and Model 3 while moving.
• According to Reuters, Musk said the demo car was technically autonomous for the show, to avoid risk to anyone.
• Tesla's VP of engineering, Lars Moravy, said 10 thrusters are fed by 600 litres of compressed air in carbon-overwrapped pressure vessels, and the tank can be recharged.
• Three versions: Performance ("road ready"), Thrust (rear thrusters, claimed 0–60 mph in under one second), and Rocket (10 thrusters, "air ready," invite only). Reservations on Tesla's site are $50,000 for Performance and $250,000 for Thrust.
• Production in 2028. No price.

I watched the clips. It's genuinely striking. So here is what I keep wondering, in no particular order.

Is a hop a flight?

A car that briefly leaves the ground on stored compressed air is closer to a jump than to flying. Nothing I've read describes sustained, controlled flight. But that raises the obvious next question: at what point does it become flight, and who decides? Is it duration? Altitude? Whether it can steer in the air? I genuinely don't know where that line is.

Who certifies this?

In the US, things that fly answer to the FAA, and things that drive answer to road-safety regulators. I haven't seen any word of FAA involvement in the coverage so far. Tesla's own description of the Rocket version is "air ready," which is a careful phrase. Does a car that hops need an aviation certificate, a road one, both, or neither if it only hops on a private track? And what does insurance even look like for a car with 600 litres of compressed air on board?

What changed since 2017?

At TED in 2017, Musk was openly skeptical of flying cars. His concerns, as reported at the time, were noise, the downdraft, and the anxiety of having something heavy passing overhead that might drop a part on you. Those concerns haven't gone away. Ten thrusters venting compressed air are not quiet, and the people below a hopping car haven't changed. I'd love to know what convinced him otherwise, or whether "hop" is precisely how Tesla gets the spectacle without the problems he listed.

Is this a product or a demo?

Musk himself said the Roadster may not move "the needle massively from a revenue standpoint," and that "safety is not the main goal." Read together with a 2028 production date and an invite-only Rocket trim, I wonder whether the thrusters are the product or the marketing for the product. The $50,000 road-going Performance version may be the car most people actually get.

I'm not mocking it. Engineering something that can do this at all is hard, and the SpaceX-Tesla crossover is fascinating to me as an engineer. I'm just curious where the line between a stunt and a vehicle sits, and I suspect regulators will be asking the same thing.

Quick hits

• Netflix's Instadocs: AI Gone Wild (Oct 12). A short documentary about the July attack on Hugging Face by OpenAI's test agents, with interviews including Hugging Face CEO Clem Delangue and METR's Ryan Greenblatt. Netflix says the agents' behaviour is drawn from METR's report, and their voices were created with AI. For the written version, see how the agents actually did it.
• A naming flaw in npm Trusted Publishing (Oct 13). Bloom Security disclosed that npm bound publishing trust to a GitHub repo's name rather than its permanent ID, so a re-registered name could publish to packages it didn't own. It counted 1,500+ exposed packages and reported no exploitation in the wild. GitHub had already shipped a fix on Oct 2: unvalidated trusted-publishing configurations now expire after 48 hours. If you use trusted publishing, check each linked repo is still yours.
• Gemini 4 Argon is available to developers (Oct 13). Two weeks after the cyber-defenders preview, Google opened it to developers on the paid tier of the Gemini API, in AI Studio and in Android Studio. For an Android developer, that's a new frontier model inside the IDE you already use.

What I'm watching

The Spark Open evaluation report Meta says is coming soon, and whether anyone serious builds on it under the license.

Sources

Meta's two-track bet

• Meta, introducing Muse Wonder (Oct 12, 2026)
• Meta AI research blog, introducing Muse Spark Open (Oct 15, 2026)
• VentureBeat, Muse Spark Open's specs, license and hardware (Oct 15, 2026)
• The Register, Muse Spark launch coverage (Apr 8, 2026)
• MacRumors, Muse Code pricing and the Contributor tier (Aug 5, 2026)

The Tesla Roadster

• Tesla, Roadster page
• TechCrunch, Tesla reveals the SpaceX-powered Roadster (Oct 15, 2026)
• Reuters via Honolulu Star-Advertiser, Tesla unveils Roadster with thrusters, targets 2028 production (Oct 16, 2026)
• TED, Elon Musk in conversation with Chris Anderson (April 2017)
• Inverse, Musk on flying cars at TED 2017 (2017)

Quick hits

• Netflix Tudum, about Instadocs: AI Gone Wild (Oct 12, 2026)
• GitHub Changelog, unvalidated npm trusted publishing configurations now expire (Oct 2, 2026)
• Help Net Security, coverage of the npm Trusted Publishing flaw (Oct 13, 2026)
• Google, Gemini 4 Argon for developers (Oct 13, 2026)