Mira Murati, former CTO of OpenAI, has just unveiled Inkling: the first AI model from her startup Thinking Machines Lab, and it arrives with open weights and a permissive license. In a world where the most powerful models are closed, this bet on openness adds pressure on the giants and opens new doors for developers. We tell you what it is, how it compares, and what changes.

On July 15, 2026, Thinking Machines Lab —founded by Mira Murati after her departure from OpenAI— launched Inkling, its first proprietary model. What's relevant isn't just who signs it, but how: it's open-weight, under an Apache 2.0 license, which allows any developer to download, modify and adapt it. A move that goes against the closed trend of OpenAI, Anthropic and Google.

What Inkling is

According to TechCrunch, Inkling is a mixture-of-experts type system with 975 billion total parameters, although it only activates around 41 billion for each task, which makes it more efficient. It was trained on 45 trillion tokens of text, image, audio and video, and reasons natively across those four formats. It's available for download in repositories like Hugging Face.

What "open-weight" means and why it matters

An open-weight model publishes its weights (the trained parameters), so anyone can download it, run it on their own infrastructure and fine-tune it for their use case. With the Apache 2.0 license, moreover, it can be used commercially with a lot of freedom. It's the opposite of closed models, which you only access via API without seeing or modifying what's inside.

The strategic bet: customization over power

Thinking Machines makes a bet different from the rest. As Axios reported, the company believes that companies care less about having the "smartest" general model and more about having one they can make their own: customize, control and optimize for their business at a lower cost. In fact, the company itself admitted that Inkling is not the best model on the market in absolute terms: its advantage lies in adaptability.

How it compares with Llama and Mistral

Inkling enters territory already occupied by open models like Llama (Meta) and Mistral. The key difference it proposes is the combination of scale (large mixture-of-experts), native multimodality and a very permissive license, backed by a top-tier team. Competition in the open-weight space is good news for developers: more options, better performance and less dependence on a single provider.

• Efficient scale: many total parameters, few active per task.

• Native multimodality: text, image, audio and video in a single model.

• Permissive license: Apache 2.0, suitable for commercial use.

What changes for developers

For those who build products with AI, an open model of this quality means being able to run frontier AI without tying themselves to a closed API, with more control over costs, privacy and customization. It's an invitation to learn to work with open models, do fine-tuning and deploy them. If you want to take the first step, check out our guide on how to learn artificial intelligence from scratch, and for the assisted-development angle, what vibe coding is and how to program with AI.

Recommended Coderhouse course

Working with open models is an increasingly sought-after skill. These training programs prepare you according to your level:

• To understand the landscape: the Introduction to Artificial Intelligence Course gives you the basics.

• To build with AI: the AI Engineering Course goes deeper into how to deploy and adapt models.

• To create autonomous agents: the AI Agents Course teaches you to take models to applications that execute tasks.

Start today: the open-weight ecosystem is growing fast; training now puts you ahead of the curve.

Frequently asked questions

What is Inkling, Mira Murati's model?

It's the first AI model from Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati. It's open-weight, multimodal, of mixture-of-experts architecture, and was launched under an Apache 2.0 license on July 15, 2026.

What does it mean for a model to be open-weight?

That its trained parameters (weights) are available to download, run on your own infrastructure and adjust to your use case. It's the opposite of a closed model, which you only access via API without being able to modify it.

Is Inkling better than the models from OpenAI or Google?

The company itself acknowledged that it's not the best model on the market in absolute terms. Its value proposition isn't to be the most powerful, but the most adaptable: one that companies can customize and control at a lower cost.

How does Inkling compare with Llama and Mistral?

It competes in the same space of open models, differentiating itself by its large scale with mixture-of-experts architecture, its native multimodality and a very permissive license, backed by the Thinking Machines Lab team.