Mira Murati, the former technical director of OpenAI, took a decisive step. Her company, Thinking Machines Lab, introduced the first AI model called Inkling. The presentation took place on July 15, 2000. The new neural network has received fully open weights for free use. The model supports multimodality and a huge context window of up to a million tokens. It is already available for further training through a special Tinker platform. The developers have published the full weights of the model on the popular Hugging Face platform. This launch was the first major product of the young company. It has significantly increased competition in the open large language models market. The main goal of the developers is to create flexible systems to adapt to the needs of users.

Inkling architecture is based on the modern principle of Mixture-of-Experts. The overall model has nine hundred and seventy-five billion parameters. However, only forty-one billion are activated during operation. This approach ensures high speed and low computing costs. The company also introduced a compact Inkling-Small version. It has twelve billion active parameters. In a number of tests, the younger version shows results close to the flagship. Developers are not chasing absolute leadership in standard ratings. They rely on effective thinking and simple customization. Internal tests have shown excellent resource efficiency. The model reaches the Nemotron 3 Ultra performance level, but spends three times less tokens. This significantly reduces the cost of running complex agent systems.

Inkling is good at programming and logical reasoning. She scored high on the AIME math test in 2000. The model also perfectly analyzes images and audio recordings out of the box. The community quickly appreciated the importance of this launch for open artificial intelligence. Analysts noted a rare combination of a million-dollar context and multimodality. Users called the release a landmark event for the entire industry. This is the first model fully trained from scratch on the new NVIDIA systems.