Thinking Machines Lab just shipped its first model, and instead of pretending it has beaten everyone, it's betting that companies want an AI they can actually take apart and rebuild.
Mira Murati left OpenAI as chief technology officer in September 2024. She spent the next several months building a company, hiring researchers and raising money. On July 15, that quiet period ended. Thinking Machines Lab released Inkling, its first in-house model. The weights are posted openly on Hugging Face. All of them. You can download it, run it and inspect it without routing a single query through Thinking Machines' servers.
Inkling is a mixture-of-experts model with 975 billion parameters in total. It activates about 41 billion of them for any given task, according to the company's announcement and reporting from The Wall Street Journal. That's the efficiency trick here: the full model is huge, but every query doesn't wake the whole thing up. Thinking Machines says it trained Inkling from scratch across text, images, audio and video, with the model built to handle those inputs directly rather than relying on a vision module bolted on afterward.
Here's the part that matters most. Thinking Machines says plainly that Inkling is not the strongest model available, closed or open. Frankly, that's an unusual thing for a lab to admit at launch. Most releases lead with a benchmark chart. This one leads with an admission.
Selling a base, not a black box
The company is pitching Inkling less as a finished product and more as raw material. It's live today for fine-tuning on Tinker, Thinking Machines' model-customization platform, letting organizations reshape the base model around their own data instead of prompting a closed API and hoping for the best. A lighter variant, Inkling-Small, is being previewed now, with its weights expected after testing wraps up.
That's a direct contrast with how OpenAI, Anthropic and Google have run their businesses. Those labs sell access to general models tuned to serve everyone at once, updated on their own schedule, with the weights locked away. Take it or leave it. Thinking Machines is arguing that a lot of enterprise customers don't want that bargain. They want a base they can own and retrain, then deploy closer to their own systems, even if it means giving up a few points on a leaderboard.
It's a bet with real precedent. Meta's Llama models, Mistral and DeepSeek have all shown that open weights can pull developers away from closed APIs, especially once a company needs to tune for a narrow domain like legal review or customer support transcripts. Ollama built an entire company on making open models easy to run locally. Thinking Machines is now trying to be both the model maker and the customization layer in one package, through Tinker, rather than leaving that work to a third party.
The money behind the model
None of this happens in a vacuum. Thinking Machines raised $2 billion in seed funding in 2025 at a $12 billion valuation, a huge round for a startup that had not yet shipped a product, according to Reuters and Bloomberg's earlier coverage of the company. Investors were betting on Murati and two of her OpenAI research colleagues, John Schulman and Lilian Weng, not on customer traction. Inkling is the first real evidence of what that bet bought.
The Wall Street Journal also reported that Thinking Machines has tied its model work to Nvidia hardware and a broader partnership announced earlier this year. That matters because open weights don't make compute cheap by magic. If you want a near-trillion-parameter model to be useful outside a research blog post, you still need infrastructure, tuning tools and enough customers willing to do the work.
Bridgewater Associates gives the pitch a concrete test case. As the Journal noted, the hedge fund used Tinker to fine-tune an open model for financial document triage and said the result beat GPT-5 and Claude Opus on that task while cutting compute costs sharply. You shouldn't read one enterprise example as proof of a market. But you can read it as the shape of Thinking Machines' argument: don't buy the single smartest model for everything, build the model that fits the work in front of you.
Whether enterprises actually want a model they have to tune themselves, rather than one that already works out of the box, is the open question here. Fine-tuning takes engineering time most companies don't have lying around. But if Thinking Machines is right that the frontier model race has started optimizing for the wrong thing, benchmark scores instead of adaptability, Inkling won't need to beat GPT or Gemini everywhere. It needs enough developers willing to build on top of it. Not rent someone else's black box.
Also read: Whatnot Acquires Shaped to Speed Up Live Shopping Recommendations • Bitcoin Miner MARA Buys Texas Power Site to Chase the AI Boom • A Caltech Startup Shrank a 27 Billion Parameter AI Model to Fit on an iPhone
