Thinking Machines Lab is nearing a $5 billion funding round led by Accel, Business Insider reported November 6, citing several people familiar with the deal. The investment would value the AI company at $40 billion before the new money and about $45 billion afterward if all $5 billion goes into the company. Thinking Machines declined to comment to Business Insider, and Accel did not respond to its request for comment.
The lab raised $2 billion in July 2025 in an early round led by Andreessen Horowitz. Accel lists itself as a seed investor. At the time of that round, founder Mira Murati, the former chief technology officer of OpenAI, had yet to release the lab’s first product. Now developers can pay Thinking Machines to adapt an AI model to a specific job without running their own fleet of chips.
What developers use Tinker for
A team using Tinker brings its own training examples and decides how to score the model’s answers. Thinking Machines runs the training on its chips and saves versions of the model as it learns. Glean, which sells search software for workplaces, says it used Tinker to adapt a model that plans searches. It decides which queries to run and when to stop. Glean’s own search system still finds the documents. Glean says its setup cuts response time by roughly half and uses about 25% fewer tokens, the pieces of text a model processes, compared with its previous approach.
That service now includes Inkling 2 Small, which Thinking Machines introduced November 5. The company says the model uses about 17 billion of its 594 billion parameters at a time. Parameters are values the model learns during training. The design lets different parts of the model handle different inputs, while the model draws on a small share of its full size at a time.
Thinking Machines used the new model to show what Tinker could do on a forecasting job. It gave Inkling 2 Small a $500 Tinker budget and asked it to adapt a smaller Qwen model to predict values in a series of measurements. The lab says the run cost $426 on Tinker and took a little over four hours. When the researchers tested it on data set aside from training, the adapted model made about 15% less forecasting error than the original Qwen model. Thinking Machines made Inkling 2 Small available through Tinker first. It plans to release the model’s weights, the values developers need to run their own copies, under an Apache 2.0 license after outside safety testing.
Nvidia and Thinking Machines plan a larger buildout
In March, Nvidia and Thinking Machines announced a multiyear partnership to deploy at least one gigawatt of Nvidia’s Vera Rubin computing systems for model training and customization. They targeted early 2027 for the first deployment. Nvidia also said it had invested in Thinking Machines and gave no dollar amount in the announcement.
The Information separately reported in September that Nvidia was in talks to put about $2.5 billion into a proposed $5 billion to $6 billion round. That report covered negotiations at the time. Business Insider named Accel as the lead investor in November, and the companies have yet to disclose the full investor list or say how Thinking Machines would spend the money.
