Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, has officially unveiled its first open-source AI model — Inkling. This is not just another neural network, but an ambitious attempt to redefine standards in the open model segment, betting on multimodality and flexibility.
Inkling's architecture is built on the Mixture of Experts (MoE) principle. This means that only part of the network is activated when processing each request, dramatically accelerating inference without losing analytical depth. The total number of parameters is 975 billion, of which 41 billion are active. The context window reaches 1 million tokens — allowing the model to process massive data arrays, including long documents and complex dialogues. Training was conducted from scratch on a dataset of 45 trillion tokens, covering text, images, audio, and video.
Inkling's key feature is its ability to reason simultaneously across multiple modalities (text, image, audio), although it currently only generates text-based outputs, including code and structured data. Users can independently adjust the balance between speed and quality by configuring the reasoning level. This makes the model ideal for tasks requiring either a quick response or deep analytics.
Leadership in Agentic Tasks, but Vulnerable to China
Inkling demonstrates outstanding results in benchmarks focused on agentic behavior. On the MCP Atlas test, which measures the percentage of completed tasks using the Model Context Protocol, Inkling scored 74.1% — nearly 30 points higher than Nvidia Nemotron 3 Ultra, the main Western competitor among open models. On SWE-Bench Verified (autonomous bug fixing in GitHub), the result was 77.6% versus 70.7% for Nemotron. In the FORTRESS Adversarial test, which evaluates correct handling of malicious requests, the model scored 78%.
However, compared to Chinese counterparts, Inkling's position appears less confident. On Terminal Bench 2.1, Inkling scored 63.8%, while Z.ai GLM 5.2 achieved 82.7%, and the recently released Kimi K3 from Moonshot AI reached 88.3%. The latter also leads in the PhD-level scientific reasoning test (Humanity's Last Exam).
Thinking Machines positions Inkling as a "universal" model that does not sacrifice performance in some tasks for superiority in others. The bet is on customization: the company believes that businesses need a flexible model tailored to a specific task rather than a monolithic solution. Alongside the main version, a lightweight Inkling-Small (276 billion parameters, 12 billion active) has been introduced, but its full weights have not yet been published — the model is still undergoing testing.
My expert opinion: Inkling is a powerful signal that the future of AI lies not with closed giants like GPT-5, but with open, customizable systems. However, the gap with Chinese models in coding and scientific tasks shows that Western developers have catching up to do. Inkling's success will depend on how quickly the community can adapt it to their specific niches.