Thinking Machines, the artificial intelligence (AI) startup founded by former OpenAI Chief Technology Officer Mira Murati, announced Wednesday the release of Inkling, a massive open-weight AI model.
The launch marks a bold bid to provide Western enterprises with a viable alternative to dominant open-source models currently coming out of Chinese AI labs.
Inkling enters a market where the Western open-source ecosystem has lagged Chinese competitors like Alibaba’s Qwen. The gap widened after Meta Platforms Inc. pivoted toward a more proprietary approach following the release of Llama 4.
To bridge this divide, Inkling is designed as a mixture-of-experts (MoE) system containing 975 billion parameters, though it routes only 41 billion active parameters per task to optimize speed and cost.
Inkling is trained on 45 trillion tokens of text, image, audio, and video, allowing native reasoning across all four mediums. It features adjustable “thinking effort” controls, allowing developers to trade speed for accuracy, and actively flags uncertainty rather than hallucinating.
In early tests, the model achieved comparable coding performance to NVIDIA Corp.’s Nemotron 3 Ultra while using only a third of the tokens.
“The open-weight model market has been a Chinese story for over a year. Inkling gives Western enterprises a credible alternative positioned on customization economics, shifting spend from per-token API pricing to infrastructure the enterprise controls,” said Mitch Ashley, vice president and practice lead for Software Lifecycle Engineering and AI-Native Software Engineering at The Futurum Group. “Engineering teams should treat base-model selection as an architecture decision. The model an organization fine-tunes becomes part of its software substrate and switching costs compound with every downstream customization. That evaluation cannot be deferred.”
While Thinking Machines acknowledged Inkling is not the absolute strongest model on the market, the startup is betting on customizability. Rather than offering a rigid, general-purpose chatbot, the company is positioning Inkling as a base model for organizations to fine-tune themselves.
This strategy is already yielding results. In a collaborative project with Bridgewater Associates, researchers used Thinking Machines’ customization platform, Tinker, to fine-tune an open model on specialized financial data. The resulting model scored 84.7% on financial reasoning benchmarks — outperforming top closed-source alternatives at approximately one-fourteenth of the operational cost.
Thinking Machines brought Inkling to market in just nine months, a fraction of the multi-year development timelines seen at OpenAI and Anthropic. The startup trained the model on Nvidia’s advanced GB300 NVL72 systems under a strategic compute partnership.
Rather than charging metered API fees for model access, Thinking Machines plans to generate revenue through Tinker, charging for training, fine-tuning, and the surrounding hosting ecosystem.
With a headcount now stabilized at roughly 200 employees, the startup’s debut marks a critical test of whether highly customizable, open-weight AI can successfully disrupt the lucrative proprietary monopolies of Silicon Valley’s tech giants.
