Thinking Machines Lab shipped its first in-house AI model Wednesday. It’s called Inkling, and according to TechCrunch AI, it’s open-weight, meaning anyone can download it and modify it directly. That’s the opposite of what OpenAI, Anthropic, and Google do with their flagships.
The startup, founded by former OpenAI CTO Mira Murati, spent about a year and a half building infrastructure quietly. This is the first thing the public can actually touch.
What Inkling Actually Is
The specs, per TechCrunch AI:
- 975 billion total parameters, but it’s a mixture-of-experts system that only activates about 41 billion for any given task. Standard trick for keeping huge models fast and affordable.
- Trained on 45 trillion tokens across text, image, audio, and video. It reasons natively across all four.
- Text-only outputs for now, including code, styled artifacts, and structured data.
- Calibrated answers. It flags uncertainty instead of guessing.
- Adjustable “thinking effort.” Dial it down when you want speed over depth.
On one coding benchmark, Thinking Machines says Inkling hits the same performance as Nvidia’s Nemotron 3 Ultra using a third as many tokens.
The Unusual Part: They Say It’s Not the Best
Most launches come with a chart showing the new model on top. This one doesn’t. Thinking Machines’ own blog post says Inkling is “not the strongest overall model available today, open or closed.”
That’s not modesty. It’s positioning. The company isn’t selling Inkling as a finished product. It’s selling it as a starting point that organizations fine-tune themselves through Tinker, its model-customization platform.
What stands out here is the strategic bet underneath. Thinking Machines argues that AI trained centrally by one lab and then frozen underperforms AI that organizations shape around their own expertise. Because the expertise that matters lives with the people who hold it, not in a general-purpose chatbot.
The Proof Point
The strongest evidence comes from a project with Bridgewater Associates, the world’s largest hedge fund. Researchers took an open-source model and trained it further on Bridgewater’s financial expertise.
Result: 84.7% on financial reasoning tests, beating top proprietary models, at roughly a fourteenth of the running cost.
Worth noting: that evaluation came from Bridgewater and Thinking Machines, not an independent third party. Take it accordingly.
Why the Timing Works
Thinking Machines isn’t arguing this alone anymore. Microsoft CEO Satya Nadella published a post Sunday warning that enterprises on proprietary models pay twice: once in subscription fees, and again by handing over business knowledge baked into their prompts and corrections, which gets absorbed into future model versions. Microsoft has billions invested in both OpenAI and Anthropic, which makes that a notable thing to say out loud.
Hugging Face CEO Clem Delangue told TechCrunch AI something similar last week. Frontier models get reserved for experimentation and high-value work. Most production AI shifts to private or open-source alternatives. That’s exactly the split Thinking Machines is building around.
The Caveats
A few things worth holding onto:
- Fine-tuning isn’t free. It takes serious machine learning talent. And customers, not Thinking Machines, own responsibility for making their customizations safe.
- Partial distillation. Inkling was pre-trained from scratch, but the company says it used other open-weight models, including Moonshot AI’s Kimi K2.5, to generate some early post-training data. It insists the next model will use fully self-contained post-training.
- The money question is open. Thinking Machines struck a March partnership with Nvidia for a gigawatt of Vera Rubin capacity and trained Inkling entirely on GB300 NVL72 systems. It hasn’t said how it’ll pay for that. A reported $50 billion round was coming together last November, then stalled by January. The company won’t discuss funding.
What Comes Next
Thinking Machines is emphasizing speed. OpenAI took roughly five years to reach market and show revenue. Anthropic took about three. Thinking Machines says it did it in nine months.
The real question isn’t whether Inkling beats GPT or Claude on a leaderboard. It’s whether enough enterprises have the talent and the appetite to customize their own models rather than rent someone else’s. If Nadella and Delangue are right about where production AI is heading, Thinking Machines timed this well.
If they’re wrong, an open-weight model that admits it isn’t the strongest is a hard sell.
Full details are available at the original source.