Thinking Machines Lab is reportedly seeking $5 billion to $6 billion. The company sells researchers and software developers a way to customize AI models without operating the computer clusters that train them. Its Tinker service leaves customers in control of their training data and algorithms while handling the infrastructure behind the work.
A team using Tinker chooses a supported starting model and provides examples or an environment in which it can learn. The developer writes the training process on a local computer, while remote graphics processors perform the intensive calculations. The service lets the team test outputs and save progress. This separates the question of what a model should learn from the engineering job of keeping a large training system running.
A concrete use case is improving search inside a business. Thinking Machines describes how Glean used Tinker to train a smaller model to select useful search tools and gather documents for a larger reasoning model. That is a specialized job inside a broader application, rather than an attempt to replace every task with one general chatbot. The example illustrates why companies with their own useful data may want to adapt a model themselves.
The commercial product is usage-based training access, with a separate charge for stored checkpoints—the saved state of a model during training. Thinking Machines says customers can download those checkpoints and that it does not use their training data to train its own models. Its documentation offers worked examples for coding, mathematical reasoning and tool use, giving developers starting points while leaving them responsible for evaluating the resulting behavior.
MT Newswires reported September 4, citing The Information, that the proposed raise would value the business at at least $40 billion before the new investment and that Nvidia could supply roughly half the money. The Next Web's September 6 account describes the larger range as an update to the earlier report and says Accel is in talks to lead. These remain reported discussions, not a company-announced closing, and the terms can change.
The product's test is whether customers get enough value from repeated training experiments to keep paying for them. Tinker removes some infrastructure work, but useful customization still needs suitable examples and a way to judge whether the trained model performs better. More specialized search, coding and research workloads would give the service a continuing role beyond a single model launch.
Correction and update, September 8: Captables previously carried a $1 billion proposed raise. Updated source reporting describes $5 billion–$6 billion. The financing ledger now uses $5 billion, the bottom of that reported range, replacing the earlier entry rather than adding a second round.