Thinking Machines Lab
AI research company building multimodal systems for real-time human-AI collaboration.
Founded
2025AI research company building multimodal systems for real-time human-AI collaboration.
The company's products center on Tinker, a training API for fine-tuning open-source language models that exposes four primitives: forward_backward, optim_step, sample, and save_state. It supports LoRA fine-tuning across a broad catalog of open models.
In July 2026 Thinking Machines Lab released Inkling, its first open-weights model: a mixture-of-experts with 975 billion total and 41 billion active parameters, a 1-million-token context window, and native text, image, and audio modalities. Weights are distributed through Hugging Face.
In July 2025 the company raised a $2 billion seed round at a $12 billion valuation led by Andreessen Horowitz, with participation from Nvidia, Accel, ServiceNow, Cisco, AMD, and Jane Street. That makes it one of the best-funded AI startups of its cohort.
Multibillion-dollar compute agreements back its next generation of models: a gigawatt-scale NVIDIA Vera Rubin partnership announced in March 2026 and a Google Cloud deal in April 2026. These commitments point to continued frontier-scale training and inference investment.
The company pairs frontier-model research with production training infrastructure, so teams can fine-tune open-weights models, including its own Inkling, through a single API. Its research bench includes authors of widely used systems such as PyTorch and Segment Anything.
Tinker's usage-based pricing and checkpoint storage make experimentation cost-predictable, and Inkling's open weights let customers deploy and adapt the model on their own infrastructure instead of depending on a proprietary API.
As a company founded in 2025, Thinking Machines Lab fields a smaller research organization than the established frontier labs and a narrower product surface, currently one training API and one open-weights model family.
Inkling enters a crowded open-weights market where larger labs release competing models with deeper compute and distribution advantages, and the company's training capacity depends on third-party GPU supply secured through large infrastructure partnerships.
Tinker bills on usage, measured in US dollars per million tokens, with checkpoint storage charged separately at $0.10 per gigabyte-month. The model ties cost directly to experimentation volume rather than seats or subscriptions.
Inkling is distributed as open weights at no license cost, so monetization runs through Tinker usage and infrastructure partnerships, including the company's compute agreements with NVIDIA and Google Cloud.