Inkling by Thinking Machines: An Open-Weight AI Model You Run on Your Data — What It Means for Business

On July 15, 2026, Thinking Machines — the company led by Mira Murati, formerly OpenAI's chief technology officer — released its first AI model, called Inkling: an open-weight model that anyone can download, run, and modify. The direct answer for a business owner: Inkling's significance is not that it is the strongest — the company itself says it is not the strongest overall — but that it is a capable model whose copy you can own, run on your own servers so your data never leaves, and customize on your company's data for better performance at lower cost. That is exactly what matters to a Saudi business that wants private, cost-controlled AI.

What is Inkling, briefly?

Inkling uses a Mixture-of-Experts architecture: 975 billion total parameters, but it activates only about 41 billion of them for any given task, which makes it lighter and cheaper to run than its size suggests. It was trained on 45 trillion tokens of text, images, audio, and video, and understands all four modalities natively rather than through bolt-on add-ons. It supports a context window of up to 1 million tokens — enough for whole documents and contracts in a single request — and its full weights are available to download on Hugging Face, while it can be customized on the company's Tinker platform. Thinking Machines is also previewing a lighter version, Inkling-Small, with 276 billion total parameters, for those who want lower running costs.

"Open-weight" is the real news for your business

Most leading models (from OpenAI, Anthropic, and Google) are closed: you send your data to the provider's server and pay per use. An open-weight model, by contrast, gives you the model file itself, so you run it inside your own infrastructure or in a cloud within the Kingdom. Why does this matter? Because your data — customer contracts, records, trade secrets — stays with you and does not go to a third party, which simplifies your compliance with the Personal Data Protection Law overseen by the Saudi Data and AI Authority (SDAIA). It also gives you independence from a single vendor: no sudden price hikes, and no service shutdown that halts your operations overnight.

Customizing on your data = higher performance at lower cost

Thinking Machines' message is clear: it bet on customization rather than a "one size fits all" model. Through the Tinker platform you can fine-tune Inkling on your domain's data — your customers' language, your documents, your policies — so that on your specific task it becomes more accurate than a larger general model, and at lower running cost, because you run a model tuned to your need instead of calling the biggest model on the market for every small request. This logic — a smaller, customized model outperforming a giant general one locally — is the core practical value for most companies, not the race to top the benchmark charts.

Multimodal: text, image, and audio in one model

Because Inkling understands text, image, audio, and video natively, its practical uses widen: analyzing an invoice or a scanned document, transcribing a customer's voice note and grasping its intent, reviewing a product photo, or extracting data from a PDF contract. Instead of wiring together several separate tools for each input type, a single model handles all these modalities — which simplifies building systems on it and reduces points of failure.

But "open weights" is not really free

Honesty matters here: downloading the model at no charge does not mean running it is free. Operating a model this size requires powerful hardware (expensive GPUs) and operational expertise to tune, monitor, and secure it. The company itself also acknowledges that Inkling is not the strongest overall compared with the latest closed models. So for intermittent use, a fast start with an idea, or the hardest general tasks, a ready-made closed API may remain simpler and more suitable. The right decision is not "open is always better," but "which is better for your case."

When should you choose an open model like Inkling?

Favor the open model when these come together: sensitive or regulated data that should not leave the organization; large, steady usage that makes per-request billing costly over time; a need for deep customization to your domain and your customers' language; and internal operational capability or a technical partner to manage the deployment. But if you are starting out or testing an idea, begin with a ready-made API, then move to open weights once the need matures, the volume grows, and the savings and privacy justify the investment.

How to benefit from Inkling with Origami

At Origami we treat open models as an engineering tool, not a shiny headline: we first assess whether Inkling even fits your case versus a closed alternative, then — if it is the better fit — we deploy it in a private environment inside your infrastructure, fine-tune it on your data, and integrate it with your systems (an internal assistant over your documents, automation, or multimodal input processing), with security and governance controls that preserve your compliance. The goal is a measurable practical outcome, not adopting technology for its own sake.

Conclusion

Inkling's launch reinforces a trend that serves companies: powerful open models you can own, run privately, and customize on your own data. It is not the optimal choice for every case, but it opens a real door to private, cost-controlled AI that is independent of a single vendor — which suits many Saudi organizations that want to benefit from AI without giving up their data.

Sources

• Thinking Machines Lab — official Inkling announcement: https://thinkingmachines.ai/news/introducing-inkling/
• Hugging Face — Inkling model page and weights: https://huggingface.co/blog/thinkingmachines-inkling
• Launch coverage (July 15, 2026): TechCrunch and Axios. Figures and specifications are as stated in the official sources as of the publication date.