Thinking Machines Labs releases open-weight AI model Inkling with 975B parameters and native multimodal training

Thinking Machines Labs, an AI startup founded by former OpenAI CTO Mira Murati, publicly released its first proprietary AI model, Inkling, positioning it as an open-weight alternative to proprietary systems from OpenAI, Anthropic, and Google.

Inkling is a mixture-of-experts architecture with 975 billion total parameters but uses approximately 41 billion active parameters per task, a design intended to maintain speed and cost efficiency at scale.

The model was trained on 45 trillion tokens spanning text, image, audio, and video, enabling native multimodal reasoning, though its current generation capabilities are limited to text output such as code, styled artifacts, and structured data.

Thinking Machines emphasizes Inkling’s calibrated outputs, including explicit uncertainty flagging, and a user-adjustable "thinking effort" setting to trade off speed and reasoning depth.

The company claims Inkling achieves comparable coding performance using roughly one-third the tokens of Nvidia’s Nemotron 3 Ultra, based on internal benchmarking.

Thinking Machines explicitly states Inkling is "not the strongest model available today, closed or open," focusing instead on well-rounded performance and customizability for enterprise use cases.

The model is intended as a starting point for organizations to fine-tune via Tinker, the company’s model-customization platform, requiring significant machine learning expertise for safe deployment.

Thinking Machines argues that centrally trained, one-size-fits-all models underperform systems customized by domain experts, citing a collaboration with Bridgewater Associates where a fine-tuned open-source model reportedly scored 84.7% on financial reasoning tests at a fraction of the cost of proprietary systems.