Mira Murati’s Thinking Machines Lab has released Inkling, a new multimodal open-source AI model trained from scratch. The model’s weights are available for free download under an Apache 2.0 license on Hugging Face, positioning it as a Western alternative for developers.
Inkling is designed as a generalist model and offers a capable open-weights option for developers concerned about using models developed in Beijing. This release follows Murati’s departure from OpenAI in September 2024. Thinking Machines Lab was founded in February 2025 and reportedly raised $2 billion in July 2025 at a $12 billion valuation. However, reports in November 2025 about seeking a new round at a $50 billion valuation collapsed by January 2026.
Technical Specifications and Capabilities
Inkling is a mixture-of-experts model featuring 975 billion total parameters, with 41 billion active parameters per task. It processes text, image, and audio inputs and supports a context window of 1 million tokens. The model was pretrained on 45 trillion tokens, encompassing text, images, audio, and video.
Performance Benchmarks
Data indicates Inkling scored 74.1% on the MCP Atlas benchmark, nearly 30 points higher than Nvidia’s Nemotron 3 Ultra (44.2%). On the SWE-Bench Verified benchmark, Inkling achieved 77.6%, surpassing Nemotron 3 Ultra’s 70.7%. Inkling also recorded 78.0% on FORTRESS Adversarial, reportedly the highest score among all compared open-weights models.
However, analysis suggests Inkling does not lead in all benchmark categories. Z.ai’s GLM 5.2 scored 82.7% on Terminal Bench 2.1, compared to Inkling’s 63.8%. Kimi K2.6 is noted as leading on the Humanity’s Last Exam benchmark. According to the FactSheet, Inkling is not the strongest model available today, whether open or closed.
Significance as a Western Alternative
The open-source nature of Inkling is highlighted as a key advantage for Western developers hesitant to use models developed in Beijing due to legal, security, or compliance reasons. The availability of its weights under an Apache 2.0 license allows for free download and self-hosting, offering a viable alternative.
Thinking Machines Lab also previewed Inkling-Small, a version with 276 billion total parameters and 12 billion active parameters. The exact release timeline for Inkling-Small weights remains uncertain.
Overall, Inkling represents a significant development in the Western open-source AI model landscape, offering advanced multimodal capabilities and a competitive alternative. Its performance relative to leading Asian models continues to be a point of comparison.
Broader Context
Source materials place the factual news in this context: Mira Murati left OpenAI in September 2024.
Leave a Reply