The Shift Toward Enterprise-Specific AI

Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, has released its first model, Inkling. Unlike the general-purpose models from major labs, Inkling is an open-weight, mixture-of-experts (MoE) system. It features 975 billion total parameters, utilizing 41 billion per task to balance performance with operational efficiency. The company argues that centralized, one-size-fits-all models are inherently limited because they cannot capture the domain-specific expertise unique to individual organizations.

Customization Over Centralization

The core strategy behind Inkling is to provide a foundation that enterprises can fine-tune using the company’s 'Tinker' platform. This approach addresses two major enterprise concerns: the high cost of subscription-based proprietary models and the risk of leaking proprietary business data through prompts and feedback loops. By allowing companies to own and customize their models, Thinking Machines aims to deliver higher value for specialized tasks. Evidence for this approach includes a recent collaboration with Bridgewater Associates, where a fine-tuned open-source model outperformed top proprietary models on financial reasoning tests while costing roughly 1/14th as much to run.

Technical and Economic Trade-offs

Inkling is designed for flexibility, allowing users to adjust 'thinking effort' to trade off between speed and accuracy. While the company admits Inkling is not the 'strongest' model available, it prioritizes well-rounded performance and efficiency. Economically, the company is betting that its revenue will stem from its customization platform (Tinker) and ecosystem support rather than metered API access. This model represents a departure from the high-burn, centralized-compute-heavy strategies of competitors like OpenAI and Anthropic, focusing instead on enabling organizations to build and host their own specialized AI infrastructure.