What Thinking Machines Lab does
Thinking Machines Lab (thinkingmachines.ai) is an AI research and product company building (1) open-weight frontier-style multimodal models and (2) a managed training platform intended to let researchers and teams customize model behavior by controlling training and fine-tuning rather than relying only on prompt-time techniques. The company positions its core mission as making AI systems more widely understood, customizable, and “generally capable,” with safety as a first-order release constraint and with a focus on human-AI collaboration.
More
From a product standpoint, Thinking Machines has two main public surfaces: - Inkling and Inkling-Small are open-weights models described as reasoning-oriented multimodal systems with variable “thinking effort” and support for long context and tool-oriented use. The company also reports that both models can be fine-tuned on its platform. - Tinker is a managed training API for researchers and developers. It abstracts away infrastructure concerns while exposing primitives for training control (e.g., forward/backward passes, optimization steps, sampling, and checkpoint saving). Tinker is described as supporting LoRA fine-tuning workflows, using internal clusters to handle scheduling, resource allocation, and failure recovery. Strategically, Thinking Machines is operating at the intersection of the open-weights model wave and the tooling wave for customization. Its differentiation is not only that it releases open models, but that it couples those releases with a training workflow designed to help downstream teams iterate quickly on model behavior (including safety and capability-evaluation loops). The company’s approach also emphasizes safety testing as part of the open-weight release process. In terms of traction and near-term direction, the company’s own newsroom shows a rapid product cadence in 2025–2026: Tinker general availability and upgrades (including OpenAI-API-compatible sampling and vision input support), initial open-weight model releases, and ongoing research-previews (e.g., “interaction models”) alongside grants programs designed to distribute compute in support of safety, teaching, and research.