Murati's Thinking Machines Releases Its First AI Model
N.R. Finch
Thinking Machines, founded by former OpenAI CTO Mira Murati, has released its first foundation model Inkling with fully open weights. The pitch is not benchmark dominance — it is enterprise customization on your own data, deployed on your own infrastructure — a direct challenge to the closed-source pricing of OpenAI and Anthropic.
What exactly is Inkling?
Inkling is an AI foundation model trained from scratch by Thinking Machines. Full weights are live on Hugging Face, with fine-tuning available on the company's own platform, Tinker.
The company says upfront: Inkling is not the most powerful model available today. This means → it is not competing for the top of general-capability leaderboards. The bet is that enterprise customers care more about owning and reshaping the model than about raw benchmark scores.
In plain terms = others sell "the smartest brain." Inkling sells "a brain you can rewire and take home."
Why go open instead of closed?
Thinking Machines' core thesis is clear: enterprises ultimately want AI they can own, not just rent from a leaderboard winner.
Palantir CEO Alex Karp recently told CNBC that closed-source frontier AI tools are overpriced and opaque on IP protection — echoing the same market read.
This reflects a widening crack in the enterprise AI market — a growing segment of corporate buyers feel they are paying OpenAI and Anthropic without getting enough control in return.
How does Murati herself view "openness"?
Murati was at OpenAI when the full GPT-2 weights were withheld in 2019 over misuse concerns — the moment OpenAI began retreating from its open-source origins.
This means → her stance on openness is not absolute: open when risks are manageable, closed when they are not.
Inkling being open does not guarantee future models will follow. The company has already flagged that stronger models are in training, and a lighter Inkling-Small will be released after testing.
Whose data and compute powered the training?
Inkling was trained from scratch, not fine-tuned from any third-party model. However, the final training stage used data generated by existing open models — including Kimi K2.5 from China's Moonshot AI (月之暗面).
Training ran on Nvidia's latest AI infrastructure — Nvidia is also an investor in Thinking Machines.
This reflects the standard topology of AI startups today: compute depends on Nvidia, and data ecosystems cross borders. Open models are already feeding each other in a loop.
A $12 billion valuation — does this release deliver?
Thinking Machines closed a $2 billion seed round at a $12 billion valuation in 2025 — with no model or product shipped at the time, a record.
The company has since reportedly signed a multi-billion-dollar Google Cloud partnership.
Inkling is the first public test of whether that enormous funding can translate into real enterprise competitiveness through the open-weight customization playbook. The verdict is only beginning to take shape.
Content is for reference only, not financial advice.