Member-only story
The launch demo wasn’t a benchmark chart. It was a terminal session in which Inkling — the new 975B-parameter model from Mira Murati’s Thinking Machines Lab — wrote its own fine-tuning job, ran it through the company’s Tinker API, evaluated the result, and loaded its own new weights back into the coding harness. The target behavior: become a lipogram model that never uses the letter “e,” something prompting alone can’t reliably achieve. Seven steps later, the model was answering questions with no e’s — because it had post-trained itself.
Within hours of the July 15 release, Artificial Analysis scored Inkling at 41 on its Intelligence Index — making it the top-scoring open-weight model ever released by a US lab, three points above NVIDIA’s Nemotron 3 Ultra (38) and far above Gemma 4 31B (29) and gpt-oss-120b (24).
And here’s the part almost nobody saw coming: the launch post says, in plain text, “Inkling is not the strongest overall model available today, open or closed.” A frontier lab spent 18 months and part of a $2 billion seed round training a model from scratch, then led with a confession.
I spent the weekend reading the technical report, digging through the Hacker News thread, and running the model through Tinker’s playground and…