Thinking Machines Open Sources “Inkling,” Challenging Centralized AI Giants

In a major bid to challenge the dominance of “one-size-fits-all” proprietary artificial intelligence, the highly capitalized startup Thinking Machines Lab has officially launched its first in-house AI model, Inkling. Founded by former OpenAI CTO Mira Murati, the San Francisco-based “neo lab” released the open-weight model on Wednesday, July 15, 2026, marking its first major public proof point after nearly a year and a half of quiet infrastructure building. Inkling is designed from scratch as a massive, natively multimodal Mixture-of-Experts (MoE) system that processes text, image, and audio inputs. By making the model’s weights completely public under an enterprise-friendly Apache 2.0 license, Thinking Machines is positioning itself directly against the closed, cloud-only systems of tech giants like OpenAI and Anthropic.

Technically, Inkling boasts 975 billion total parameters, though its sparse MoE architecture activates only about 41 billion parameters per token to keep operational costs and latency low. It has been trained on a massive dataset of 45 trillion multimodal tokens. Although the model’s current output is restricted to text and structured data, it is designed with extreme flexibility in mind. Rather than chasing raw benchmark supremacy, Thinking Machines has built Inkling as a sturdy base model that developers can download, host locally, and tailor to specific tasks via the company’s Tinker optimization platform. This launch addresses a critical “why” for the industry: as rising cloud API costs and vendor lock-in force global corporations to reconsider their AI budgets, customizable open-weight models allow businesses to maintain sovereignty over their data and lower their inference bills.

This paradigm shift has quickly garnered industry backing. On day zero, major cloud and development platforms like Databricks and Modal announced direct integrations. According to the Databricks Blog, Inkling is now accessible through its Unity AI Gateway, allowing developers to run coding workflows and build internal agents securely within their own database perimeters. Simultaneously, developer platform Modal highlighted that their custom speculation engine can run Inkling up to 67% faster, delivering highly interactive agentic workflows.

The debut of Inkling sits in a broader lineup of major technology and AI developments that unfolded on July 15, 2026. The shift toward specialized utility is keeping the tech industry incredibly active. For instance, TechCrunch reported that Spotify founder Daniel Ek secured a staggering $700 million funding round to scale his body-scanning and preventative medicine startup, Neko Health, highlighting a parallel trend of using specialized tech for highly specific, real-world workloads. Meanwhile, centralized giants are feeling the pressure of both legal battles and physical constraints. Amidst an ongoing legal dispute with Apple over hardware trade theft, OpenAI unexpectedly released a $230 light-up mechanical keyboard designed to integrate with its Codex agentic coding application. At the same time, Microsoft showcased the practical applications of its security-first AI tools, patching a record-breaking 570 software vulnerabilities during its monthly Patch Tuesday cycle.

By delivering a highly capable open-weight competitor, Thinking Machines has made a definitive statement on the future of AI. Rather than routing all global intelligence through a handful of closed cloud portals, they are betting that the next wave of innovation will belong to companies running, tweaking, and securing their own private models. With weights now hosted on Hugging Face and fine-tuning tools active on Tinker, Inkling offers a highly customizable foundation to prove that theory.