Inkling open-weight AI model targets enterprise customization over general-purpose benchmarks

Thinking Machines Lab released Inkling, an open-weight AI model that organizations can download and customize via the Tinker platform. Unlike the flagship models from OpenAI, Anthropic, or Google, Inkling is designed as a starting point for enterprise-specific adaptation rather than a finished general-purpose system.

What happened

Inkling uses a mixture-of-experts design with 975 billion total parameters but draws on about 41 billion for any given task. It was trained on 45 trillion tokens across text, image, audio, and video, with native reasoning across all four modalities. Current generation capabilities are limited to text output, including code, styled artifacts, and structured data.

The model is designed to give calibrated answers, flagging uncertainty rather than guessing, and lets users adjust thinking effort to trade accuracy for speed. Thinking Machines does not claim Inkling is best-in-class. Its briefing materials state explicitly that Inkling is "not the strongest model available today, closed or open."

Full weights are available on Hugging Face and the model is live for fine-tuning on Tinker, the company's customization platform. Thinking Machines used data from existing open models, including Moonshot AI's Kimi K2.5, in its final training phase. The company says it pre-trained Inkling from scratch but used other open-weight models to generate early post-training data before large-scale reinforcement learning took over.

A project with Bridgewater Associates reportedly scored 84.7% on financial reasoning tests, beating top proprietary AI models while costing roughly a fourteenth as much to run. Those results come from the two companies' own evaluation, not an independent one.

Why AI builders should care

Inkling represents a bet that enterprises want AI they can customize rather than rent from a handful of frontier labs. The company argues that centralized labs underperform compared to organization-specific expertise because so much knowledge is specific to the people who hold it.

Microsoft CEO Satya Nadella recently warned that enterprises using proprietary AI models effectively pay twice: once in subscription costs, and again by handing over business knowledge embedded in prompts and corrections, which can be absorbed into future model versions. Hugging Face CEO Clem Delangue predicted that frontier models will increasingly be reserved for experimentation while most production AI work shifts to private or open-source alternatives.

For AI builders, Inkling offers a path to own the model weights and fine-tune on proprietary data without per-token licensing. The revenue model for Thinking Machines is Tinker, not the model itself, via training, fine-tuning, and hosting ecosystem partnerships.

Practical implications

If you are building AI products for enterprise customers, Inkling changes the calculus in a few ways:

• You can download the weights and fine-tune on your own data without sending data to a third-party API. This matters for regulated industries and companies with strict data governance requirements.
• The mixture-of-experts design keeps inference costs lower than dense models of similar total parameter count. The company claims Inkling uses a third as many tokens as Nvidia's Nemotron 3 Ultra to hit the same coding performance.
• You need serious ML talent to customize effectively. Fine-tuning a 975B-parameter model is not trivial. Thinking Machines is betting that enterprises will pay for Tinker's tooling rather than doing it from scratch.
• The calibrated uncertainty feature is useful for agentic workflows where you want the model to signal when it doesn't know rather than hallucinate. Combined with adjustable thinking effort, this gives builders more control over latency and reliability.

Caveats

Inkling is not a frontier model. The company is transparent that it does not beat closed models on general benchmarks. Its value proposition is customization, not raw capability.

The training used distillation from other open-weight models, which may raise questions about originality and long-term differentiation. The company says the next model will use fully self-contained post-training.

Public disclosures on funding remain limited. A reported $50 billion fundraising round was said to be stalling earlier this year. The company has declined to discuss its funding picture since, though Nvidia made a significant investment when the companies announced a partnership for Vera Rubin compute.

Finally, the Bridgewater results are self-reported and not independently verified. The model's performance on financial reasoning may not generalize to other domains without similar customization effort.