Thinking Machines Lab Releases Inkling, Their First Open-Source AI Model

Inkling was just released on July 15, 2026 by Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati. It's a significant release because it's the company's first in-house AI model, and unlike the flagship models from OpenAI, Anthropic, or Google, it's open-weight, meaning outside developers and companies can download it and modify it directly.

This is, hands down, the largest open-source AI model that any American AI lab has made, which makes it even more important, as China is slowly taking over the open-source AI market.

But, is it any good?

First, let's talk about the platforms you can get access to this model through. There's Tinker, their first-party cloud service, where you can get access to the model through an API. Thinking Machines Lab also offers fine-tuning for the model through this platform. You can also use third-party providers, like Together AI, to run the model. And because the model is open-source, if you have the hardware, you can download and run it using Hugging Face. If you just want to try it out, the Inkling Playground in the Tinker console lets you chat with the model for free for a limited time.

Now, the specs.

Inkling is a massive model with 975 billion parameters in total, but it doesn't use all of them at once. Thanks to its Mixture-of-Experts design, it only activates about 41 billion parameters for any given task. That's what keeps it fast and affordable to run despite its size.

It also has a context window of up to 1 million tokens, but when served with Tinker, it comes in only 64K and 256K context window variants.

Under the hood, it's a 66-layer transformer. Instead of one big neural network doing all the work, it splits tasks across 256 specialized expert sub-networks (plus 2 that are always on), but only 6 of those experts get used per token. Think of it like having a huge team of specialists on staff, but only pulling in the handful you actually need for each specific job.

It also handles attention (how the model tracks relationships between words) in a clever way, mixing local attention that focuses on nearby context with global attention that can look across the entire conversation, at a 5:1 ratio. This combo is likely part of why it can handle that massive 1-million-token context window without completely blowing up compute costs.

So, is it actually good? In the Artificial Analysis Intelligence Index, it gets a 41, which is right under MiMo V2.5 Pro and right over DeepSeek V4 Flash. This isn't the most promising performance, considering that this model costs roughly $1.10/M Input Tokens and $4.00/M Output Tokens, which is about the same price as GLM 5.2, a model with a score of 51.

In terms of output token efficiency, it also isn't the best, taking around 25k tokens per task, making it on par with MiniMax M3 and Mistral 3.5 Medium.

But raw benchmarks might not be the point here. Thinking Machines isn't trying to build the smartest model on earth. Their bet is that most companies don't need the smartest model on earth. They need a good model that they can shape to fit their exact use case. That's where Tinker comes in. The model is free. The fine-tuning platform is the product.

And there's already evidence this works. In a collaboration with Bridgewater Associates, researchers used Tinker to fine-tune an open model with specialized financial data. The result was a lightweight model that scored 84.7% on leading financial reasoning benchmarks, outperforming the most advanced proprietary alternatives at less than 10% of the cost.

Inkling is, as the name suggests, just the beginning of a new era of American open-source AI. It might not be the best model on earth by a long shot, and it doesn't need to be. It's trying to be the best starting point, and that might matter more.