Mira Murati Just Shipped an Open Model — Inkling and the Datacenter Catch
Mira Murati's Thinking Machines shipped Inkling — the strongest open-weight model a US lab has released, free to download and nearly impossible to run. It takes 488 gigabytes just to load, which quietly moves the real barrier from the license to the hardware.
The Promise
The Risk
What actually shipped
On July 15th, 2026, Mira Murati’s Thinking Machines Lab released its first model. It’s called Inkling, and the specs are real: a mixture of experts with 975 billion total parameters, about 41 billion of them active for any given task, trained on 45 trillion tokens across text, image, audio, and video, with a one-million-token context window. Artificial Analysis scored it 41 on its intelligence index — the leading open-weights model from a US lab, and Thinking Machines is candid that it trails the best Chinese open models. The license is Apache 2.0. Download it, change it, ship it, no strings.
That combination is why the launch landed as a milestone. It’s also why the coverage mostly missed the part that matters.
Where the pitch meets the wall
Sparse activation is the promise. The router lights up only 41 billion parameters per answer, so you pay a mid-size compute bill for frontier-scale knowledge. But sparse activation lowers your compute cost, not your memory cost. The router can reach any of the 975 billion parameters at any moment, so all of them have to be loaded and resident, all the time. Squeeze every weight to four bits and Inkling still needs roughly 488 gigabytes of memory just to hold — before the context window, before a single user connects.
488 gigabytes is a datacenter number. Not an RTX 5090. Not a 256GB Mac Studio. Both are a no-go.
What “open” actually buys
I’ve watched this pattern for 25 years in cybersecurity. “You can read the source” was never the same as “you can secure it.” Open weights are walking the same road: they transfer ownership of the model, not independence from compute. The barrier didn’t disappear when the license went permissive — it moved from the license to the hardware. And Apache 2.0 with no guardrails hands every safety and governance decision to whoever deploys it.
For an organization with a hard data-residency or fine-tuning requirement, that trade may be worth it. For everyone else, the honest read is simpler: a frontier model in everyone’s hands turns out to mean a frontier model in the hands of anyone who owns a datacenter. That’s a much shorter list.