Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati, released its first in-house model on Wednesday morning: Inkling. The model’s notable feature is that it is open-weight, allowing outside developers and companies to download and modify its weights directly.
Technical details and capabilities
Inkling is a 975 billion-parameter mixture-of-experts (MoE) system, but it typically uses only a fraction of that—about 41 billion parameters—for any given task. That design is common for very large models because it reduces latency and operating cost. According to the company, Inkling was trained on 45 trillion tokens spanning text, image, audio, and video and is able to reason natively across those modalities. For now, however, its outputs are text-only, including code, stylized artifacts, and structured data.
The company says Inkling provides calibrated answers, flagging uncertainty rather than guessing, and allows users to adjust a “thinking effort” setting to trade off speed against thoroughness. On one benchmark cited by Thinking Machines, Inkling uses roughly one-third as many tokens as Nvidia’s Nemotron 3 Ultra to reach comparable coding performance.
Strategy: customization and rapid go-to-market
Inkling is Thinking Machines’ first public demonstration after roughly a year and a half of building AI infrastructure largely out of public view. The company previously previewed “interaction models” in May—systems intended to listen, speak, and even interrupt rather than merely wait for user input like typical chatbots.
The startup’s central thesis is that AI organizations can adapt themselves will outperform one-size-fits-all models trained centrally by a single company. Inkling is positioned less as a finished product and more as a starting point for customers to fine-tune using Thinking Machines’ model-customization platform, Tinker. That approach also places responsibility for safe customization on customers; fine-tuning requires significant machine-learning expertise.
Performance, partnerships, and cost transparency
Thinking Machines does not claim Inkling is the strongest model available. Its briefing material explicitly states Inkling is “not the strongest model available today, closed or open.” The company emphasizes balanced performance and customizability instead.
Training ran on Nvidia hardware after a March partnership that secured a gigawatt of Vera Rubin compute capacity; Inkling was trained entirely on Nvidia GB300 NVL72 systems. Thinking Machines has been guarded about costs and revenue strategy. Public reporting indicates revenue has not been the immediate priority and that a reported $50 billion fundraising round discussed last November had stalled by January. The company has not provided further detail on fundraising since.
Use of other models and the distillation question
Asked whether Inkling was trained on outputs from competitors (a practice known as distillation), the company says it pretrained the model from scratch but did use other open-weight models—including Moonshot AI’s Kimi K2.5—to generate some early post-training data before large-scale reinforcement learning was applied. Thinking Machines says the next model will use fully self-contained post-training.
Case work and evidence for customization advantages
Thinking Machines points to a recent collaboration with Bridgewater Associates, the world’s largest hedge fund (which is not a Thinking Machines investor), as evidence for its approach. In that project, researchers further trained an open-source model on Bridgewater’s financial expertise; the combined system reportedly scored 84.7% on financial reasoning tests and cost roughly one fourteenth as much to run versus top proprietary models. Those results come from the two companies’ internal evaluation rather than an independent benchmark.
Headcount and company culture
Thinking Machines now employs roughly 200 people, recovering from a wave of departures earlier in the year that included two co-founders who left for OpenAI in January. According to a source inside the company, its culture is intentionally built to favor continuity and avoid dependence on single personalities, which the company views as reducing the impact of staffing changes.
What this means for the market
Thinking Machines is betting that offering an open-weight model plus a customization platform will attract enterprises that want to keep and refine their own institutional knowledge, instead of sending it to centrally trained proprietary models. Industry figures such as Microsoft CEO Satya Nadella and Hugging Face CEO Clem Delangue have recently voiced views consistent with this direction, suggesting more production workloads may migrate to private or open-source models while frontier models remain for experimentation and high-value tasks.
Inkling is available as a downloadable starting point for organizations that have the expertise and governance capacity to fine-tune and host it. For Thinking Machines, the business model appears to focus on Tinker-based customization services and the hosting ecosystem around a public-weight model rather than metered inference fees from model access.
