Skip to content

AI

Mira Murati's Inkling: 975B parameters, 41B doing the work

Thinking Machines Lab ships its first model, Inkling, and hands out the weights. Most of its 975 billion parameters nap while about 41 billion handle each task.

Cover photo: Michael Gaylard from Horsham, UK / CC BY 4.0

TL;DR

  1. Inkling is open-weight: 975 billion parameters total, about 41 billion active per task.
  2. Trained on 45 trillion tokens of text, image, audio and video; outputs text only for now.
  3. Thinking Machines Lab, founded by ex-OpenAI CTO Mira Murati, says it reasons across all four modalities.

What happened

Thinking Machines Lab, the startup founded by ex-OpenAI CTO Mira Murati, has released its first in-house model. It is called Inkling, and it landed on a Wednesday morning.

The headline spec is size. Under the hood, Inkling uses a mixture-of-experts design with 975 billion parameters in total. For any single task, it wakes up only about 41 billion of them. That is a common design trick for keeping giant models quicker and cheaper to run.

The company says Inkling reasons natively across text, image, audio and video. What comes out is narrower for now: text only, covering code, styled artifacts and structured data. So it can take in the whole sensory buffet, but it answers in plain typing. Image and sound generation are not on the menu yet.

The model also arrives after a long stretch of low-profile infrastructure work. A May research preview showed "interaction models", AI built to listen and speak, even interrupt, rather than sit and wait like a typical chatbot. Inkling is the first public proof point of that effort.

Why it matters

The flagship models from OpenAI, Anthropic and Google are closed. Inkling is open-weight, so outside developers and companies can download it and modify it directly. That is a different deal from renting an API and hoping the terms stay friendly.

A well-funded startup with a famous founder choosing open weights gives developers one more serious option at the top end of the size chart. More choice means more pressure on everyone charging by the token.

Caveats, because we have standards. The size, the training data and the multimodal reasoning all come from the company's own release materials. We have no independent benchmark numbers here, so treat "how good is it" as an open question until people start poking at it.

Open weights mean you can inspect and modify the thing instead of just renting it.

Mixture-of-experts also matters in practical terms. A model that only uses about 41 billion parameters per task is a very different thing to serve than one that uses all 975 billion every time.

What to do with it

  • Check the weights.If you run models yourself, find the official release and see what hardware the full 975 billion parameters demand before you get excited.
  • Test on your own work.Run Inkling against your real prompts, code and documents, because vendor claims are not your workload.
  • Watch the output limits.It reasons over images, audio and video but only answers in text, so plan pipelines around that.
  • Track the interaction models.The May preview of AI that can interrupt hints at where Thinking Machines is heading, so keep an eye on it.

Keep up, sapiens. The open-weight robots just got a very large new roommate.