Thinking Machines Lab Launches 'Inkling' AI Model

Thinking Machines Lab, the startup founded by former OpenAI executive Mira Murati, has introduced its first in-house AI model named Inkling. The model features an open-weight architecture and is trained on 45 trillion tokens of multimodal data, supporting a one-million-token context window. While utilizing a massive 975 billion parameter base, the model uses a mixture-of-experts (MoE) design, employing 41 billion active parameters per task to prioritize speed and efficiency. The company released the model under an open-source license, allowing developers to host and modify it via the Tinker platform.

Inkling is designed to provide calibrated, instruction-following responses while resisting censorship. Unlike closed-source models from competitors like OpenAI and Anthropic, Thinking Machines Lab is betting on enterprise-level customization. The startup aims to generate revenue through its hosting ecosystem and customisation tools rather than direct metered access. The company partnered with Nvidia to train the model using GB300 NVL72 systems. While acknowledging that Inkling is not currently the most powerful model available, the company emphasizes its versatility for specific organizational workflows and fine-tuning needs.

The move follows recent industry shifts, including White House-mandated restrictions on new frontier AI models. As Microsoft reportedly trains sales teams to emphasize its own platform over OpenAI and Anthropic, startups like Thinking Machines Lab are positioning open-weight models as a way to provide enterprises with more autonomy. Inkling is supported by a lighter-weight version, Inkling-Small, aimed at reducing latency. As the venture moves forward, it plans to iterate on its post-training datasets to increase independence from third-party generation sources, positioning itself as a challenger to established US-based AI laboratories.