Thinking Machines Lab is no longer just a rumor.
The company, built by executives and researchers who walked out of OpenAI, has dropped its first major model: Inkling.
And they are handing out the keys.
Built From Scratch
Inkling is an open-weight model. This isn’t just another fine-tuned variant. It was trained from scratch to process audio, video, and text simultaneously.
The specs are massive. We are talking 975 billion parameters. You cannot run this on a laptop. You need a cluster of specialized chips just to get it off the ground.
So, how good is it?
It does not break the records on popular benchmarks. But that’s not the point. Inkling performs solidly on a wide range of tasks, showing off advanced reasoning and coding abilities without needing a gold-plated dataset to shine.
The Ghost in the Machine
Here is where things get interesting.
The team used Inkling to help refine itself during training. AI building AI.
As the model processed more complex tasks, its internal “chain of thought” began to shed weight. It got lazier.
“the chain of thought became more concise… dropping grammatical overhead… leaving the final response unaffected.”
The model realized it didn’t need to write out every step to get the right answer. It started thinking faster because it stopped speaking as much. A natural optimization? Or something else entirely?
The Anti-Corp Play
Open-source models are hot right now. Not because they are technically superior, but because they are cheaper to run. You aren’t paying per token to a monopoly. You can tweak the weights. You can break the rules.
Most high-performance open weights come out of China right now. Thinking Machines says Inkling competes with them. If true, that is a bold claim for a startup.
But the money is the real story here.
Thinking Machines launched with a $12 billion valuation. It was the largest seed round in history. A lot of that capital is betting on decentralization. The founders argue AI shouldn’t be locked in black boxes by five big companies.
It should be decentralized. Anyone should be able to build on their own data.
Who’s Behind It
You recognize the names.
- Mira Murati: Former CTO (and brief CEO) of OpenAI.
- John Schulman: One of the minds behind the RLHF methods that made ChatGPT safe-ish.
- Lilian Weng: The safety and research guru at OpenAI.
They left. They raised a fortune. And now they have a product.
Previously, they showed off Tinker, a tool for fine-tuning, and some voice interaction demos. Now they have Inkling.
The competition isn’t sleeping. Anthropic, another defector-backed firm, is looking at an IPO that could hit a trillion-dollar valuation. Claude is eating the lunch of many enterprises, especially for code.
The space is getting crowded. The spending is reckless.
Thinking Machines thinks breaking the weight restrictions on models changes the game.
Maybe.
But Inkling needs serious hardware. Is open source still “democratizing” AI if only a handful of rich companies can afford to run it?
