Inkling AI: The 975B Open Model from Thinking Machines

Former OpenAI CTO Mira Murati's lab just released Inkling — a 975B-parameter open-weights multimodal model with a 1M context window, Apache 2.0 licensed. Here is what it is, how it performs, and how to run it.

How to Run Inkling → vs DeepSeek & GLM
975B
Total Parameters (41B active)
1M
Context Window
45T
Pretraining Tokens
Apache 2.0
License

What Is Inkling AI?

Inkling AI is the first open-weights model from Thinking Machines Lab — the startup founded by former OpenAI CTO Mira Murati and backed by NVIDIA. Released on July 15, 2026, Inkling is a Mixture-of-Experts transformer with 975 billion total parameters (41B active per token), a context window of up to 1 million tokens, and native understanding of text, images, and audio. The full weights are on Hugging Face under Apache 2.0.

Inkling AI was pretrained on 45 trillion tokens spanning text, images, audio, and video. Thinking Machines is refreshingly direct about positioning: Inkling is not the strongest model available, open or closed. Instead it is built to be the best base for customization — multimodal, efficient through controllable thinking effort, and available for managed fine-tuning on the lab's Tinker platform from day one. The launch demo made the point vividly: Inkling wrote and ran its own fine-tuning job.

Inkling AI is the first of a model family. Alongside it, Thinking Machines previewed Inkling-Small, a 12B-active-parameter model trained with a similar recipe for dramatically lower cost and latency. This site is an unofficial Inkling AI guide covering hardware requirements, quantized versions, comparisons with DeepSeek and GLM, and the Inkling-Small preview.

Six Things That Define Inkling AI

What separates Inkling from the wave of open models it joins.

Mira Murati's first open model

Inkling AI is the first model released by Thinking Machines Lab, the NVIDIA-backed startup founded by former OpenAI CTO Mira Murati. Full weights are on Hugging Face under Apache 2.0.

Natively multimodal

Inkling reasons over text, images, and audio in one model — pretrained on 45 trillion tokens spanning text, images, audio, and video.

Controllable thinking effort

You can dial reasoning depth up or down per request, balancing cost against quality — a native feature, not a prompt hack.

Built for customization

Thinking Machines positions Inkling as a base for fine-tuning, available day one on its Tinker platform. In the launch demo, Inkling fine-tuned itself.

A model family

Inkling is the first of a family: Inkling-Small, a 12B-active-parameter preview trained with the same recipe, targets lower cost and latency.

Not chasing the crown

Thinking Machines is explicit: Inkling is not the strongest model available. It bets on multimodality, efficient thinking, and fine-tuning access instead of leaderboard wins.

Inkling AI Guides

How to Run Inkling →

Hardware requirements from 270GB quants to full 1.9TB weights, Mac Studio setups, cloud GPU costs, and hosted options.

Inkling vs Other Models →

Head-to-head with DeepSeek-V4, GLM-5.2, and Qwen3.6 — parameters, context, multimodality, and which to choose.

Inkling-Small →

The 12B-active preview that brings the Inkling recipe to consumer hardware. Everything known so far.

Inkling AI FAQ

What is Inkling AI?
Inkling AI is the first open-weights model from Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati. Released July 15, 2026, it is a 975B-parameter Mixture-of-Experts multimodal model (41B active) with a 1M-token context window, licensed Apache 2.0.
Who made Inkling AI?
Thinking Machines Lab, founded by Mira Murati (former OpenAI CTO) and backed by NVIDIA among other investors. Inkling is the lab's first released model, trained from scratch on 45 trillion tokens.
Is Inkling AI free?
The weights are free under Apache 2.0 — you can download, modify, and use them commercially. Running it is the cost: full weights need ~1.9TB of storage and serious hardware, though hosted options like Tinker and Databricks let you try it without any.
What can Inkling AI do?
Inkling reasons natively over text, images, and audio, and is strong at coding, agentic tool use, RAG, chat, and multilingual tasks. It features controllable thinking effort — dialing reasoning depth up or down per request to balance cost and quality.
Is Inkling AI better than DeepSeek or GPT-5.5?
Thinking Machines says plainly that Inkling is not the strongest model available, open or closed. Its bet is different: a multimodal, efficient, Apache-licensed base that is easy to fine-tune on the Tinker platform — customization over leaderboard supremacy.
What is Inkling-Small?
A lighter-weight preview model in the same family, trained with a similar recipe but with 12B active parameters — targeting strong performance at much lower cost and latency. It makes the Inkling recipe accessible on far smaller hardware.
Where can I try Inkling AI right now?
The fastest ways are the Inkling Playground on Thinking Machines' Tinker console and the Databricks Unity AI Gateway. For local use, Unsloth publishes quantized GGUFs starting at 270GB.
What does the name Inkling mean?
An 'inkling' is a slight idea or hint of understanding. Thinking Machines has not explained the choice publicly, but it fits the lab's positioning: a starting point you shape into something more through customization.

Get Hands-On with Inkling AI

Whether you have a Mac Studio, a GPU cluster, or nothing but a browser — there is a way to run Inkling today.

How to Run It Weights on Hugging Face