Skip to content

AI Tools

A former OpenAI CTO's startup just gave away a 975 billion parameter AI model. Here is what business owners should understand before calling it free.

On Wednesday, July 15, 2026, Thinking Machines Lab, the AI startup founded by former OpenAI chief technology officer Mira Murati, released its first model trained from scratch, called Inkling. It is a Mixture-of-Experts model with 975 billion total parameters, and the company published the full weights on Hugging Face under an Apache 2.0 license, the same permissive open-source terms used for ordinary commercial software, meaning any business can download, modify, and run it without paying a license fee or asking permission. The direct answer for a business owner: this is a real, verified, and genuinely open release, not a marketing label attached to something that turns out to be restricted. But 'open' and 'free to download' describe the license, not what it takes to actually run the model. Running Inkling at its native precision takes more raw GPU memory than most small or medium businesses have ever purchased, which is the part of this story worth understanding before it changes how anyone thinks about AI costs.

By Fabio Rabelo · Founder, ATLACIS ·

What happened

Thinking Machines Lab, founded in 2024 by Mira Murati after she left OpenAI, released Inkling on July 15, 2026, its first foundation model built from scratch. According to the company's own announcement, Inkling is a Mixture-of-Experts model with 975 billion total parameters, of which about 41 billion are active for any given task, a design that keeps very large models faster and cheaper to run than their full size suggests. It supports a context window of up to 1 million tokens and was pretrained on 45 trillion tokens of text, images, audio, and video, though TechCrunch confirms its outputs are currently limited to text. The company also previewed a smaller sibling, Inkling-Small, with 276 billion total parameters and 12 billion active, aimed at lower cost and latency, with weights to follow once testing is complete. The full Inkling weights are already live on Hugging Face under an Apache 2.0 license, a permissive open-source license that allows commercial use without royalties or restrictive terms, according to VentureBeat. Thinking Machines is also making the model available today through its own fine-tuning platform, Tinker, at a limited-time 50% discount, and is extending access to third-party inference providers including Together AI, Fireworks, Modal, Databricks, and Baseten, per The Register. Notably, Thinking Machines states plainly in its own release notes that Inkling is 'not the most performant model available today, closed or open.' The company is positioning it as a flexible, customizable foundation rather than a benchmark leader.

Why it matters for business owners

Set the specific benchmark scores aside. What matters here is a real, growing appetite among businesses for AI they can adapt to their own workflow instead of renting a fixed model from a handful of frontier labs. Thinking Machines' entire strategy is built around that bet, and there is already real evidence it works: according to Reuters, the hedge fund Bridgewater Associates used Thinking Machines' Tinker platform to build a customized version of Alibaba's Qwen open model that it said outperformed top proprietary models at lower cost. Inkling gives that same customization approach a credible, Western, Nvidia-partnered option, in a space that has mostly been led by Chinese labs such as DeepSeek, Moonshot AI, and Alibaba. For a business owner, the significance is not that Inkling is the best model available. It is not, and the company says so directly. The significance is that customizable, license-free AI is becoming a real category with real vendors and real infrastructure behind it, which means it is worth understanding on its own terms rather than dismissing it as a lab experiment.

What owners should not misunderstand

The easiest mistake here is treating 'open weight' and 'free' as the same thing as 'accessible.' They are not. According to The Register, running Inkling at its native 16-bit precision requires more than two terabytes of GPU memory, roughly eight of Nvidia's B300 accelerators or sixteen H200s, hardware that costs well into the hundreds of thousands of dollars and that almost no small or medium business owns or would casually rent. A quantized version, released alongside the full weights, can run on about half that GPU count, which is still a serious data center commitment, not something a business runs on office hardware. The second detail worth separating out: Thinking Machines itself says Inkling is not the strongest model available, open or closed. A business chasing the largest parameter count in a headline is chasing the wrong number. The company's own framing is that Inkling's value is flexibility and a genuinely open license, not category-leading capability. Those are two different kinds of value, and only one of them (capability) shows up in a benchmark chart. The third detail: for almost any business, 'using Inkling' in practice will mean going through Tinker or one of the third-party inference providers Thinking Machines named, not self-hosting the raw weights. The open license still matters even then, since it means the business is not locked to one host and can move the model elsewhere later, but it is a different kind of access than running the model on your own machine.

The operational lesson

License freedom and operational accessibility are two separate questions, and a release like Inkling answers only one of them directly. Apache 2.0 answers the legal question: can a business take these weights, modify them, and use them commercially without paying a fee or asking permission. Yes. It does not answer the practical question: can a typical business actually run this model today, on its own infrastructure, at a cost that makes sense. For Inkling specifically, the honest answer is no, unless the business already operates or plans to lease serious GPU infrastructure. What a business gets in the near term is a choice between a hosted version (through Tinker or a third-party provider, priced and metered like any other AI API) and, for organizations with real infrastructure budgets, a self-hosted deployment. Both are legitimate. Neither is free in the way 'free to download' implies.

What a serious business should do next

Before treating any new open-weight model release as a cost-saving opportunity, get a real number, not the headline framing. If self-hosting looks appealing, get an actual quote for the GPU infrastructure a model of that size requires, whether that means buying hardware or leasing cloud GPU capacity, before assuming it is cheaper than an API. If hosted access through Tinker or a third-party inference provider is the realistic path, price that against what you already pay for a comparable closed model on the same workload, since 'open' does not guarantee the hosted price will be lower. Test the smaller Inkling-Small variant, once available, or the quantized checkpoint against your actual task before assuming a business needs the largest version. Whatever path a business chooses, treat the Apache 2.0 license as a long-term hedge worth having: it means the option to move the weights to a different host later stays open, even if the business starts on a hosted plan today.

The Atlacis view

Atlacis does not push one model, one vendor, or one deployment approach as the default answer. A release like Inkling is a genuine expansion of what is available, a credible, openly licensed, Nvidia-partnered alternative to the mostly Chinese open-weight field. It is also a reminder that 'open' is a licensing property, not a cost estimate. Atlacis helps owners separate those two questions before either one turns into a real budget line: what does this model's license actually let you do, and what would it actually cost, in hardware or in hosted API pricing, to run it on your real workload, sized to what your business needs rather than to the biggest number in the announcement.

The short version

  • On July 15, 2026, Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, released Inkling, a 975 billion parameter (41 billion active) Mixture-of-Experts model, with full weights free on Hugging Face under a permissive Apache 2.0 license.
  • The company states directly that Inkling is 'not the most performant model available today, closed or open.' Its value is licensing freedom and customization, not benchmark leadership.
  • Running Inkling at native precision requires more than two terabytes of GPU memory, roughly eight Nvidia B300 accelerators or sixteen H200s, hardware almost no small or medium business owns.
  • For most businesses, using Inkling in practice will mean a hosted option through Tinker or a third-party inference provider (Together AI, Fireworks, Modal, Databricks, Baseten), not self-hosting the raw weights.
  • Real business demand for customizable open models already exists: Bridgewater Associates used Thinking Machines' Tinker platform to build a custom version of Alibaba's Qwen model that it said beat proprietary models at lower cost.
  • The operational lesson is to separate two questions before budgeting for any open-weight release: what the license actually allows, and what it would actually cost, in hardware or hosted pricing, to run on a real workload.
Tags:open source AImodel selectionAI hardwareprivate AIAI vendorsAI buying decisionsbusiness AIAI decision-makingAI infrastructurevendor dependency
FAQ

Common questions

Does Inkling being open-weight and free mean my business can run it at no cost?
No. The Apache 2.0 license means there is no fee to download, modify, or use the weights commercially. Actually running the full model takes more than two terabytes of GPU memory, hardware most small or medium businesses do not own. Most businesses would access Inkling through a hosted plan (Thinking Machines' own Tinker platform or a third-party inference provider), which is priced like any other AI API, not free.
Is Inkling the best AI model a business could use right now?
Thinking Machines says directly that it is not the most performant model available, open or closed. Its stated advantage is flexibility for customization and a genuinely open license, not top benchmark scores. Whether it fits a specific business depends on the task, not the parameter count.
What should a business actually check before considering an open-weight model like Inkling?
Get a real cost figure for however you would actually access it: a hosted API quote if using Tinker or a third-party provider, or a GPU infrastructure quote if self-hosting. Test a smaller variant against your real task before assuming you need the largest version. Confirm what the open license actually buys you, mainly the ability to move the model to a different host later, since that is a real, separate value from raw capability.

Make better AI decisions, starting with one call.

Book a free AI Fit Call. We will tell you what to use, what to avoid, and where to start. No jargon, no pressure.