Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati, announced its first in‑house model Inkling on Wednesday morning. Unlike the flagship models from OpenAI, Anthropic or Google, Inkling is open‑weight, meaning developers and companies can download the weights and modify them directly.

Technical Architecture

Inkling follows a mixture‑of‑experts design with a total of 975 billion parameters, yet only about 41 billion are engaged for any given task—a pattern that keeps ultra‑large models faster and cheaper to run. Trained on 45 trillion tokens spanning text, images, audio and video, the model claims native reasoning across all four modalities, although its current output is limited to text, code, styled artifacts and structured data.

Open‑Source vs. Closed‑Box Paradigm

The startup’s core thesis is that organizations that can tailor AI to their own data and workflows will outperform the one‑size‑fits‑all models sold by the biggest labs. To that end, Inkling is marketed less as a finished product and more as a starting point, with the company’s Tinker platform enabling enterprises to fine‑tune the model themselves. This shifts the burden of safety and compliance onto the customer—a point echoed by Microsoft CEO Satya Nadella’s warning that proprietary AI forces enterprises to pay twice: once for subscriptions and again for the knowledge embedded in prompts.

Performance Benchmarks and Competitive Context

According to Thinking Machines, Inkling uses roughly one‑third the tokens of Nvidia’s Nemotron 3 Ultra to achieve comparable coding performance on a standard benchmark. While the company does not claim Inkling is the strongest model available today—open or closed—it emphasizes balanced performance and customizability. A recent collaboration with Bridgewater Associates demonstrated the model’s financial reasoning ability, scoring 84.7 % on proprietary tests while costing about one‑fourteenth of the expense of leading closed models.

Future Roadmap and Economic Viability

Inkling was pre‑trained from scratch, but the company admits it leveraged other open‑weight models such as Moonshot AI’s Kimi K2.5 for early data generation before large‑scale reinforcement learning took over. The next generation model will rely entirely on self‑contained post‑training. Inkling was trained on Nvidia’s GB300 NVL72 systems under a gigawatt‑scale Vera Rubin compute partnership, yet the firm remains tight‑lipped about monetization, having postponed a reported $50 billion fundraising round.

Speed to Market

Thinking Machines highlights its rapid development timeline—nine months from concept to release—against OpenAI’s five‑year and Anthropic’s three‑year journeys to market. If the model’s cost‑efficiency and customization promise hold, it could reshape how enterprises approach AI, moving the focus from monolithic, subscription‑driven services to adaptable, open‑weight solutions.