Murati returns with a bang: Thinking Machines Lab releases its first multi-modal open source model Mira Murati, former chief technology officer of Inkling OpenAI, recently officially announced the first major achievement of the "Thinking Machines Lab" she founded -

Murati returns with a bang: Thinking Machines Lab releases the first multi-modal open source model Inkling. Former OpenAI chief technology officer Mira Murati recently officially unveiled the first major achievement of the "Thinking Machines Lab" she founded - the multi-modal AI model Inkling. With its strong performance, this model is regarded by the industry as the strongest competitor in the open source AI field in the United States. It is reported that the founding team of Thinking Machine Laboratory has a luxurious lineup. About two-thirds of the core members are from OpenAI, and have been deeply involved in key tasks such as cutting-edge research, product implementation and security mechanism construction. The Inkling launched this time is the first model independently trained by the team. It adopts a hybrid expert (MoE) architecture, has a total parameter volume of 975B, an activation parameter of 41B, and supports a context window of up to 1M tokens. In the pre-training stage, Inkling digested massive data of 45 trillion tokens, covering multiple modalities such as text, images, audio and video. Currently, Thinking Machine Laboratory has fully disclosed the model weights on the Hugging Face platform and its own Thinker API. Through horizontal benchmark test comparison, Inkling has demonstrated excellent comprehensive strength in multiple fields. Especially in terms of mathematical solving, Inkling achieved a high score of 97.1% in the AIME2026 test, slightly higher than DeepSeek V4Pro; and in the MCP Atlas test of the agent workflow, it also far surpassed Nemotron3Ultra with a score of 74.1%. Although there is still a certain gap between Inkling and closed-source head models such as Claude Fable5 or GPT5.6Sol in reasoning, coding, and specific software engineering tasks, Inkling has shown strong competitiveness in terms of native visual and audio understanding capabilities. Especially in the audio processing benchmark test (MMAU), its score of 77.2% is very close to Gemini3.1Pro. In the current open source ecological competition, Inkling performs differently when facing strong opponents such as GLM5.2, DeepSeek V4Pro and Kimi K2.6. For example, GLM5.2 maintains its advantage in pure coding and complex reasoning tasks, while Inkling is slightly better in general instruction tracing (IFBench). Overall, as the debut of Mulati’s team, In