Former OpenAI executive Mira Murati’s startup, Thinking Machines Lab, launched its first internally developed AI model, Inkling, on July 15.
Inkling is an open-weight model trained to produce calibrated answers, follow instructions, and avoid censorship. Users can adjust the model’s reasoning effort to balance speed and depth.
Although described as a generalist system covering agentic, coding, and reasoning tasks, current outputs are limited to text, code, styled artifacts, or structured data. The company noted that this flexibility supports varied workflows beyond benchmark performance.
Unlike OpenAI, Anthropic, and Google, Thinking Machines is releasing the model under an open-source license, allowing developers to download and modify the weights from Hugging Face.
The launch occurs as U.S. support for open-source large language models grows, partly due to recent White House limits on new closed models from leading firms.
Thinking Machines positions Inkling for enterprises that want to fine-tune models themselves via its Tinker platform, rather than relying on one-size-fits-all offerings. Revenue will come from the hosting ecosystem, not direct model access.
The company now has about 200 employees after earlier departures, including two co-founders who joined OpenAI in January.
Inkling contains 975 billion parameters and supports a one-million-token context window, yet activates roughly 41 billion parameters per task through its mixture-of-experts design. It was trained from scratch on 45 trillion tokens of text, image, audio, and video.
Researchers used other open-weight models, such as Moonshot AI’s Kimi 2.5, to help create early post-training data before large-scale reinforcement learning. Future models will use fully internal datasets.
Training ran entirely on Nvidia GB300 NVL72 systems under a March partnership for gigawatt-scale Vera Rubin capacity.
On coding benchmarks, Inkling matched Nvidia’s Nemotron 3 Ultra while using one-third the tokens. A smaller 12-billion-active-parameter version, Inkling-Small, offers lower cost and latency.
The company stated that Inkling is not the strongest model available but offers a customizable open-weights base with multimodal support and fine-tuning tools.
