Thinking Machines monetizes open weights

This model works only if Thinking Machines turns free distribution into paid infrastructure and services before open weights become a commodity. The open pieces are the top of funnel. Developers can test safety tools, evals, and weights on their own, then move to hosted training and inference when they need reliable compute, easier deployment, support contracts, and custom model behavior tied to their own data and workflows.

• The clearest evidence is Tinker, which already looks like the commercial wedge. It charges per million tokens for training, supports large open weight models like Qwen and Llama, exposes low level fine tuning primitives, and lets users download checkpoints. That is a hosted developer workflow, not just a research release.
  1 sacra 2 sacra 3 thinkingmachines
• This is the same playbook used by open model companies like Mistral and Stability AI. They widen adoption with open models or permissive access, then charge for managed APIs, self hosted commercial rights, enterprise support, and deployment help. The money comes from convenience, control, and procurement readiness, not from hiding weights.
  4 sacra 5 mistral 6 sacra 7 stability
• Thinking Machines is pushing one step deeper into customization than a plain model API. Its product is built around forking a base model, uploading proprietary data, changing safety settings, and deploying managed or on premises instances. That makes enterprise support and advanced customization high value line items, because customers are buying a tailored system, not generic chat access.
  1 sacra 2 sacra 3 thinkingmachines

The next phase is a race to become the default control plane for open model customization. If Thinking Machines keeps shipping hosted tooling like Tinker, open source can pull in developers while premium infrastructure, dedicated compute, and hands on model adaptation become the durable revenue layer. That would place it closer to an AI platform vendor than a pure model lab.