Thinking Machines open weights to paid infrastructure

This model makes Thinking Machines more like an AI cloud company than a pure model lab. Open weights help spread the model and attract researchers, but the money comes when those users bring data, training code, and inference traffic into Tinker, where Thinking Machines runs the GPUs, exposes low level training primitives, and charges for the actual work done. That creates a funnel from free distribution into paid infrastructure, while still letting customers export checkpoints for supported open models.1 2 3

• Tinker is not just a chatbot endpoint. Teams bring training data and their own training loops, then use primitives like forward_backward, optim_step, sample, and save_state while Thinking Machines handles the cluster. That is closer to managed ML infrastructure than to a closed model API.1
• The tradeoff is that openness grows adoption faster than it grows hosting revenue. Downloaded weights can be run on another cloud, so Thinking Machines has to win on convenience, performance, and tooling, not exclusivity. DeepSeek follows a similar pattern, using open release to drive demand for hosted usage and ecosystem reach.1 4
• Comparable platforms show where competition lands. Fireworks won customers by giving them fast access to new open models through OpenAI style APIs, strong observability, and concurrency guarantees. Prime Intellect is building a similar loop across training and serving, which means pricing power will come from owning the workflow, not just publishing a strong model.2 3

The next step is a fight to become the default place where open models are adapted for real work. If Thinking Machines can make Tinker the easiest path from a promising base model to a tuned production system, open release becomes customer acquisition and the hosted training and inference layer becomes the durable business.1 2 3