Thinking Machines Lab Releases Inkling: A 975B-Parameter Open-Weights Multimodal MoE With 41B Active Parameters And Controllable Thinking Effort
Thinking Machines Lab released Inkling on July 15, 2026 - its first model trained entirely from scratch. The weights are available publicly under the Apache 2.0 license, meaning developers and researchers can use, modify, and redistribute the model without the restrictions that typically accompany more permissive-sounding but commercially limited open-weight releases.
Architecturally, Inkling is a Mixture-of-Experts (MoE) transformer with a total parameter count of 975 billion, though only 41 billion of those are active for any given forward pass. This is a standard MoE tradeoff: the full parameter space enables broad capability and specialization, while sparse activation keeps inference costs closer to those of a much smaller dense model. The model supports native multimodal input - text, images, and audio - and carries a 1 million-token context window, which places it among the longer-context models currently available in the open-weights space.
What makes Inkling's positioning somewhat unusual is the lab's own candor about where it sits in the landscape. Thinking Machines Lab states directly that Inkling is not the strongest model available, whether compared to leading open-weights releases or closed commercial systems. The intended use case is not out-of-the-box performance on benchmarks, but rather serving as a foundation for fine-tuning and adaptation. The practical differentiator the lab highlights is controllable thinking effort - the ability to dial reasoning depth up or down depending on the complexity of a task. This kind of inference-time flexibility is increasingly relevant as developers look to balance output quality against latency and compute cost in production settings.
For teams that want a capable multimodal base they can adapt freely, Inkling offers a relatively rare combination: a large MoE architecture, broad modality support, a long context window, and no licensing friction. Whether its reasoning controllability proves useful in practice will likely depend on how well that mechanism holds up across diverse fine-tuning scenarios - something the broader research community will be able to evaluate now that the weights are publicly available.
