Thinking Machines Lab Releases Inkling Open-Weight AI Model

Thinking Machines Lab, a San Francisco-based startup founded by former OpenAI executives, released its first general-purpose foundation model, Inkling, on July 15, 2026. Thinking Machines Lab designed Inkling as an open-weight, mixture-of-experts transformer model with 975 billion total parameters and 41 billion active parameters. The model was trained on 45 trillion tokens using Nvidia GB300 NVL72 systems, supported by a multi-year partnership and significant investment from Nvidia.

Inkling natively processes text, image, audio, and video inputs with a context window of up to 1 million tokens. While the company acknowledges it is not the strongest model currently available—trailing DeepSeek V4 Pro and GLM 5.2 in some benchmarks—it outperforms Nvidia Nemotron 3 Ultra on SWE-Bench Verified and offers controllable reasoning and resistance to censorship. To address the high GPU infrastructure costs of the full model, the company also previewed Inkling-Small, a lighter version intended to reduce latency and cost for enterprises.

Thinking Machines Lab does not monetize the model itself. Instead, it generates revenue through Tinker, a paid API for fine-tuning and customization. This strategy aims to decentralize AI control and reduce reliance on proprietary APIs. Analysts noted that the model is particularly suited for domain-specific workloads and provides a Western-developed option for organizations that prefer to deploy AI on their own infrastructure.