Former Chief Technology Officer of OpenAI, Mira Murati, has officially unveiled the first major achievement of her newly established "Thinking Machines Lab" - the multimodal AI model Inkling. With strong performance, the model is regarded as the strongest competitor in the U.S. open-source AI field at present.

It is reported that the founding team of Thinking Machines Lab is formidable, with about two-thirds of the core members coming from OpenAI, and they have been deeply involved in key work such as cutting-edge research, product implementation, and security mechanism construction. The released Inkling is the first model trained by the team, which uses a mixture of experts (MoE) architecture, with a total parameter count of 975B, activated parameters of 41B, and supports a context window of up to 1M tokens.

image.png

In the pre-training phase, Inkling processed an enormous amount of data, including text, images, audio, and video. Currently, the Thinking Machines Lab has fully released the model weights on the Hugging Face platform and its own Thinker API.

Through horizontal benchmark testing, Inkling has demonstrated outstanding comprehensive capabilities in multiple fields. Especially in mathematical problem solving, Inkling achieved a high score of 97.1% in the AIME2026 test, slightly higher than DeepSeek V4Pro; while in the MCP Atlas test for intelligent agent workflow, it far exceeded Nemotron3Ultra with a score of 74.1%.

Although in reasoning, coding, and specific software engineering tasks, Inkling still has some gaps compared to closed-source top models such as Claude Fable5 or GPT5.6Sol, it shows strong competitiveness in native visual and audio understanding. Particularly in the audio processing benchmark test (MMAU), its score of 77.2% is very close to Gemini3.1Pro.

In the current open-source ecosystem competition, Inkling shows different performances against strong opponents such as GLM5.2, DeepSeek V4Pro, and Kimi K2.6. For example, GLM5.2 maintains an advantage in pure coding and complex reasoning tasks, while Inkling slightly outperforms in general instruction tracking (IFBench).

Overall, as the debut of Murati's team, Inkling provides the open-source community with a base model with powerful native multimodal capabilities. With the release of weights, this model is expected to play an important role in future AI development and applications.