AI Startup Thinking Machines Lab Releases First Open-Weight Model, Inkling

Thinking Machines Lab is positioning Inkling-Small as an open-weight option for companies seeking to run multimodal artificial intelligence on their own infrastructure. The model combines text, image and audio reasoning in a single architecture, potentially allowing businesses to keep sensitive information within private environments, customize deployments and reduce their reliance on closed, cloud-based application programming interfaces.

Thinking Machines Lab released Inkling-Small on Aug. 3, 2026. The mixture-of-experts model has 276 billion total parameters but activates 12 billion during inference, a design intended to limit computing requirements. It supports a 1 million-token context window and is available under the Apache 2.0 license. With quantization, the company says the model can run on a single NVIDIA B300 GPU, lowering the hardware threshold for self-hosted and private enterprise deployments.

Former OpenAI Chief Technology Officer Mira Murati’s team has introduced Inkling, a 975-billion-parameter model built on a mixture-of-experts architecture. The release marks a significant model debut for Murati since leaving OpenAI and gives Western developers a new entrant in an open-weight ecosystem increasingly shaped by Chinese AI labs. Its scale also puts Inkling among the largest models made broadly available for independent use and evaluation.

As of July 27, 2026, Inkling’s weights are available for download through Hugging Face under the permissive Apache 2.0 license. Early testing, however, has challenged claims that it is the strongest Western open-source model: reviewers found that it trailed a much smaller model capable of running on a smartphone in some evaluations. The results underscore that headline parameter counts — even Inkling’s 975 billion — do not necessarily translate into superior real-world performance.