Meta's two-Mac model makes per-token pricing a habit you should test, then probably break

Meta has released an 816-billion-parameter model that it says fits on two Mac Studios. I expected to spend this column on the Macs. The more interesting thing in the announcement is the list of seven companies that will sell you the same model by the token: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. I think that list does more damage to the closed labs’ pricing than any desktop demo.

First, the mechanism, because Meta’s hardware claim is better engineered than most launch-day boasts. Muse Spark Open is a mixture-of-experts model with 816B total parameters and 42B active per token, and a 1M-token context window. You need memory for all 816B, but each generated token only reads the 42B that the router picks. Apple’s M5 Ultra lets its integrated GPU address up to 512 GB per Mac Studio, so two machines hold the whole thing, and 1.2 TB/s of memory bandwidth is what decides how quickly that active slice can be streamed through for each token. Meta calls the pair a supercomputer with a terabyte of unified memory. That is marketing, but the arithmetic underneath it is sound: sparse activation is what turns a datacentre-sized model into a desk-sized one.

Then Meta stops. The post gives no tokens per second for the Mac setup, no price for the machines and no per-token price from any of the seven hosts. For a launch pitched at people who would otherwise pay closed labs per token, those are the only figures that matter, and they are missing. My guess is that Meta knows the desktop throughput is the weak spot and chose to show you Ice Cubes bug fixes and a DaVinci Resolve edit instead.

The kernel numbers show the same habit. Meta ships custom Helion GPU kernels in SGLang with 1.03 to 3.95 times kernel-level speedups on NVIDIA and AMD GPUs. The end-to-end figure on AMD MI350 is a 3% speedup. Both numbers are probably true, but an inference bill moves with the 3%, not with the kernel figures.

So why do I still think this undercuts per-token pricing from closed labs? Because of the hosts. When one lab owns the weights, its price is whatever its margin targets say. When seven inference companies serve identical open weights, they compete on the cost of running them, and that competition pushes the price towards the price of compute. The Mac pair is the ceiling on how much any host can charge you: past a certain monthly bill, you buy the hardware and the marginal token costs electricity. Most teams will never buy the Macs, but a host that charges much more than the hardware would cost over a year will lose its largest customers to it.

The objection I take most seriously is quality. A closed model that finishes a refactor first time can be cheaper than an open one that needs three attempts and a human cleaning up after it, whatever the per-token rate. Meta’s own claim is narrow, too: “the most capable US open-weight model to date” says nothing about beating closed frontier models, and the benchmark table did not survive into the text I was working from. I accept all of that, and it is exactly why per-token price is the wrong unit. The unit is cost per task that passed. A model at a fraction of the token price that fails twice as often can lose; one that fails slightly more often probably wins. The launch post has no failure rates, so each team has to measure them.

There is one more catch before anyone rebuilds their stack. The weights come under the Muse Model License with “limited commercial terms”. Meta says that lets many developers use it free, which tells you some developers cannot. Read the terms before your legal team reads them for you, especially if you resell anything built on the model.

The test is cheap to run this week. Take a week of the tickets your coding agent actually handled, replay them against Muse Spark Open on whichever of the seven hosts quotes you lowest and against the closed model you pay for now, and divide each bill by the number of tasks that passed CI without a human rewrite. If the open model’s number comes out lower, keep paying per token, but stop paying the closed lab’s rate.

Prompted by Introducing Muse Spark Open, The Most Capable US Open‑Weight Model to Date, Meta AI Research.