Thinking Machines Lab

Thinking Machines Lab's Tinker announcement describes external research usage by Princeton's Goedel team for theorem proving, Stanford's Rotskoff group for chemical reasoning, Berkeley's SkyRL team for asynchronous reinforcement learning, and Redwood Research for control research. Named external workflows are distinct from the prior general product-availability observation.

Thinking Machines has made Tinker generally available and released Inkling and Inkling-Small as open-weight models, creating real external usage pathways, though broad enterprise adoption is still emerging.

Thinking Machines has made Tinker generally available and released Inkling and Inkling-Small as open-weight models, creating real external usage pathways, though broad enterprise adoption is still emerging.

Thinking Machines is simultaneously training frontier models with Inkling, developing Tinker for customizable post-training, and advancing real-time multimodal interaction research; its official research program combines frontier model training, post-training systems and multimodal human-AI interaction.

Thinking Machines trained Inkling from scratch as a 975B-parameter MoE with 41B active parameters, 1M context and native reasoning over text, images and audio; Inkling-Small offers comparable efficiency at smaller active scale.

Thinking Machines trained Inkling from scratch as a 975B-parameter MoE with 41B active parameters, 1M context and native reasoning over text, images and audio; Inkling-Small offers comparable efficiency at smaller active scale.

Thinking Machines and NVIDIA announced a multi-year strategic partnership to deploy at least one gigawatt of Vera Rubin systems, alongside significant NVIDIA investment and a frontier-research team.

Thinking Machines and NVIDIA announced a multi-year strategic partnership to deploy at least one gigawatt of Vera Rubin systems, alongside significant NVIDIA investment and a frontier-research team.

Thinking Machines Lab raised a $2B seed round at a $12B valuation and in 2026 secured a multi-year NVIDIA partnership to deploy at least one gigawatt of Vera Rubin systems, with an additional strategic NVIDIA investment supporting long-term frontier-model growth.

Tinker exposes low-level training primitives while Thinking Machines manages distributed training infrastructure; official use spans RL, fine-tuning, multi-agent/tool-use research, multimodal models, chemistry, forecasting, continual learning and academic research, demonstrating unusually broad portability across frontier research and ML systems roles.

Foundation-model training, multimodal reasoning, post-training, systems and frontier research skills transfer exceptionally broadly across top AI labs and model platforms.

Foundation-model training, multimodal reasoning, post-training, systems and frontier research skills transfer exceptionally broadly across top AI labs and model platforms.