Thinking Machines amps up its bet against one-size-fits-all AI with its first open model, Inkling

Details

Thinking Machines' Inkling is a mixture-of-experts model: 975 billion total parameters, but only 41 billion activate for any given query — keeping inference faster and cheaper than a dense 975B model. Trained on 45 trillion tokens spanning text, image, audio, and video, it accepts all four natively and reasons across them without separate modality encoders. One benchmark stands out in Thinking Machines' briefing materials: Inkling uses approximately one-third the tokens of Nvidia's Nemotron 3 Ultra to achieve comparable coding performance — a meaningful cost advantage if it holds up beyond benchmarks. The model also supports user-configurable thinking effort: users can dial reasoning intensity up or down to trade compute for speed. It flags uncertainty rather than guessing when confidence is low. The TechCrunch piece probes a key open question: who exactly is the customer? Thinking Machines positions Inkling as an enterprise fine-tuning base through its Tinker platform, where organizations adapt the model with their own data. Fine-tuning a 975B model requires serious ML infrastructure and talent, limiting the addressable market in the near term. Critically, fine-tuning customers are responsible for ensuring their customizations are safe — safety alignment responsibility shifts to the customer, not Thinking Machines. Thinking Machines' broader thesis — that centralized one-size-fits-all models from OpenAI, Anthropic, and Google underperform AI trained on organization-specific knowledge — is now being tested in the market rather than argued in blog posts.