This $12 billion startup finally shipped something...

Two years ago, Mira Murati did something we've all fantasized about. She woke up, decided she'd had enough of her boss, and quit her job as CTO of OpenAI with no backup plan and nothing but a vague desire to create time and space for her own exploration. If you or I did that, we'd probably end up coding HTML for

food down on Market Street. But when someone like Mira does it, A16Z shows up with a $2 billion investment into any company she decides to build. The company was eventually Thinking Machines, and until last week, it's been known for being valued at $12 billion before having actually shipped anything.

But that may have now just changed because they just released Inkling, a fully open weights model trained from scratch that can see, hear, reason, fine-tune, and potentially touch things in the real world. In today's video, we'll find out how it was built, what makes it interesting, and learn why

Thinking Machines openly admits that it's not the best model in the world, and why that's the point. It is July 20th, 2026, and you're watching The Code Report. Last year, while the rest of us were spending time getting to know our newest Soham Parikh, Mira was quietly pulling off the greatest talent heist in

Silicon Valley history. When she left OpenAI, she walked out with co-founder John Shulman, VP of research Barrett Zoph, and enough senior researchers to fill a Silicon Valley polycule. A year later, the dream team shipped Tinker, an API that lets you fine-tune open weight models without managing your own

infrastructure or losing control over the training process. When it was released, ML researchers called it cool, which I imagine was a bit underwhelming if you'd just given them enough money to buy a Costco hot dog for one out of every six people on Earth. But that may have changed last week with the release

of Inkling, a mixture of experts model with 970 billion total parameters. The trick is that it doesn't use all of them at once. Every token gets routed to a small subset of specialized experts, so only 41 billion parameters actually fire per token, giving you the most intelligence of a massive model at a

fraction of the compute. It was pre-trained on 45 trillion tokens of text, images, and audio, handles a 1 million token context window, and unlike other open models that ship with a license written by feral lawyers from Meta, the weights are Apache licensed and sitting on Hugging Face right now.

On the Trust Me Bro benchmarks, it gets mugged by Fable 5 and GPT 5.6 Soul, and lands somewhere mid-table among the open Chinese models. Now, that's not great, but the timing was even worse. One day after Inkling dropped, Moonshot announced Kimik 3, a 2.8 trillion parameter monster that, if you squint,

goes bar for bar with Fable 5, and got so popular that it even burned through their compute. And so, in a sense, the team spent 2 years cooking up the most average model in existence, but that was kind of the point. Instead of competing on raw intelligence, Inkling has a dial called thinking effort that controls how

hard the model cooks. You crank it down and you get cheap, instant answers. You crank it up and you get Ashley Tisdale's hit single from 16 years ago, with the model matching Nemotron 3 Ultra on Terminal Bench while using a third of the tokens. And when you're running an agent millions of times per day, that

adds up. And the flagship demo was pretty cool. They connected Inkling to Tinker and asked it to remove its own ability to use the letter E. From there, the model wrote its own training script, generated its own data, ran the job, then loaded its own new weights that essentially lobotomized itself. They