What is Thinking Machines’ first AI model ‘Inkling’, and how is it different from ChatGPT, Claude? | Technology News

5 min readNew DelhiUpdated: Jul 16, 2026 05:28 PM Former OpenAI executive Mira Murati’s startup, Thinking Machines Lab, on Wednesday, July 15, launched its first in-house developed AI model called Inkling.
Inkling is an open-weight AI model that is trained to deliver calibrated responses, follow instructions, and res censorship. It also lets users dial the model’s thinking effort up or down based on their preferences in a trade-off for speed.
While Thinking Machines has described Inkling as a general model trained across agentic, coding, and reasoning tasks, its outputs are currently limited to text in the form of code, styled artifacts, structured data or more. “That breadth matters for customisation and real-world use: different users need models that can adapt to very different workflows, not just excel on benchmarks,” the company said.
Thinking Machines is taking a different approach from its rival labs such as OpenAI, Anthropic, and Google to release the model under an open-source license. This means that the weights of the model are freely available for outside developers and companies to download from HuggingFace and modify them directly.
Inkling enters the market as support for open-source development of LLMs gains traction in the US, driven in part the White House’s abrupt restrictions on new cutting-edge, closed AI models released leading AI companies such as Anthropic and OpenAI.
With Inkling, the former OpenAI CTO’s venture is betting that AI models which allow enterprises to adapt for themselves will outperform the one-size-fits-all LLMs rolled out most frontier AI labs in the US. In order to cater to enterprises, Thinking Machines Lab is positioning Inkling as a model for organisations to fine-tune themselves through Tinker, the startup’s model customisation platform.

However, this could also mean that organisations would be responsible for ensuring that their fine-tuned version of Inkling is safe and secure. Unlike OpenAI and Anthropic, which charge for metered access to their models, Thinking Machines is looking to generate revenue not from the model itself but the hosting ecosystem around it via Tinker.Story continues below this ad
Thinking Machines now employs roughly 200 people, up from levels reported after a wave of departures earlier this year, including two co-founders who left for OpenAI in January.
Under the hood of Inkling
Inkling is a 975 billion parameter LLM and supports a context window of one million tokens. However, it only draws on a fraction of its total parameter count, or about 41 billion, for any given task. This is primarily due to its underlying mixture-of-experts (MoE) architecture, a common design that makes it faster and cheaper to run very large models.
Inkling has been trained Thinking Machines researchers from scratch on 45 trillion tokens of text, image, audio, and video, and reasons natively across all four, according to the company’s blog post.
In the pre-training phase, the researchers said that they used other open-weight models such as Moonshot AI’s Kimi 2.5 to help generate some of its early post-training datasets prior to carrying out large-scale reinforcement learning. The next LLM developed the company will use fully self-contained post-training data, it said.Story continues below this ad
Thanks to its partnership with Nvidia, Thinking Machines trained Inkling entirely on the chipmaker’s GB300 NVL72 systems. The two companies inked a deal in March this year to deploy a gigawatt of Vera Rubin computing capacity
Performance benchmarks
On a coding benchmark, Thinking Machines claimed that Inkling achieved scores on par with Nvidia’s Nemotron 3 Ultra while just using a third of the tokens consumed the chipmaker’s latest generation open-weight model. Inkling has also been trained to run inside a variety of coding and agent harnesses with the controllable thinking feature.

However, the company said that Inkling is “not the strongest overall model available today, open or closed.” “Instead, a combination of qualities makes it a good open-weights base for customisation: multimodal capabilities, efficient thinking, and availability on Tinker for fine-tuning. Inkling is just the start: our first release in a model family we will continue to build on,” it said.
Alongside Inkling, the company also previewed a lighter-weight model version called Inkling-Small with 12 billion active parameters that achieves strong performance with even lower cost and latency.