Thinking Machines released Inkling, an open-weight customizable multimodal model

Thinking Machines Lab has been quiet about actually shipping product for a while, so Inkling landing as the company's first open-weight model is worth pausing on. It's a multimodal mixture-of-experts system, meaning it only activates part of itself for any given request rather than firing every parameter every time. Add controllable reasoning effort and a 1M-token context window, and you get a model built less for flashy demos and more for enterprises that want to actually own and shape what they're running.

That ownership pitch is the real story here. Thinking Machines is explicitly arguing against the rent-a-one-size-fits-all-API model that has defined most of the industry's business relationships so far. Instead, the company wants businesses to customize Inkling for their own workloads rather than bolt their operations onto whatever a closed API provider decides to expose that quarter. It's a bet that control matters more to serious enterprise buyers than convenience.

On benchmarks, the early read from Deedy calls Inkling the best open-weight model outside of China, putting it ahead of Nemotron 3 Ultra and slotting it somewhere between Kimi 2.5 and Kimi 2.6. That's a specific, competitive claim, not a vague pat on the back, and it places Inkling in genuinely crowded territory where Chinese labs have been setting the pace on open weights for months.

The release also got tied to Tinker, with the suggestion that the two will pair well together, hinting Thinking Machines is thinking about an ecosystem rather than a single standalone model. Coming the same day OpenAI's hardware ambitions and Anthropic's enterprise and robotics moves dominated headlines, Inkling is a reminder that the open-weight lane hasn't gone quiet just because the chat-box giants keep grabbing attention.