Thinking Machines, the AI startup led by former OpenAI CTO Mira Murati, has published its first model, Inkling. According to the company, Inkling is an open‑weight model that supports multiple modalities — text, images, audio and video — and can produce code and structured data.
The company explicitly states that Inkling is “not the strongest overall model available today, open or closed.” Instead of competing for peak general‑purpose benchmarks, Thinking Machines markets Inkling as a starting point that customers can adapt to their own needs.
Tinker: a paid path for fine‑tuning
Thinking Machines pairs Inkling with a paid platform called Tinker that is intended for fine‑tuning and customization. The company presents Inkling as a downloadable base: the end value depends on further tuning by the user. The firm cites a joint project with hedge fund Bridgewater, in which an open model trained on Bridgewater’s domain expertise reportedly achieved 84.7% on a financial‑reasoning metric while incurring roughly one‑fourteenth of the cost of closed competitors.
Thinking Machines also says it assembled the system in about nine months.
Why this matters
The strategic premise is straightforward: if you cannot build the single most capable general model, you reposition the market so that being a customizable base matters more than absolute leaderboard performance. That approach can appeal to organizations that can inject proprietary data and domain expertise through fine‑tuning.
But that same approach creates trade‑offs. Selling Inkling as a starting point shifts many safety, alignment and domain‑specific mitigation tasks onto customers. Effective fine‑tuning typically requires in‑house machine‑learning expertise and resources, which many potential buyers lack. In practice, a freely downloadable model can end up primarily benefiting organizations that can afford to complete and secure it.
Takeaway
Thinking Machines emphasizes customization over frontier performance: Inkling is designed to be modified by users rather than to top general benchmarks. The model’s adoption will depend on how broadly organizations can access the teams and budgets needed to safely and effectively fine‑tune it, and on whether the market prizes customizable foundations more than out‑of‑the‑box maximal performance.
