Overview
- Thinking Machines released Inkling on Wednesday, July 15, 2026, as a 975-billion-parameter, open-weight model offered through its Tinker platform and other developer channels.
- Inkling uses a mixture-of-experts design that draws a small active subset of its total parameters for each task to cut inference cost and latency.
- The company published internal benchmarks showing Inkling is competitive on many tasks and notably strong on agent-style workloads, but closed models from Anthropic, Google and OpenAI still hold an overall performance edge.
- Thinking Machines says Inkling was trained from scratch on text, audio and video and that the lab used the model to fine-tune itself, though the company’s claims have not yet been independently verified and the model needs specialized chip clusters to run.
- Founded by former OpenAI leaders with large seed funding, Thinking Machines positions Inkling as a step toward decentralizing model development and lowering costs through customizable, downloadable weights, but broad adoption will depend on third-party evaluation and firms’ ability to supply the required compute.