AI Foundations Free ~7 min read New · July 2026

Inkling: what it means when a lab gives away the model.

On July 15, Thinking Machines Lab — the company founded by former OpenAI CTO Mira Murati — released its first AI model, Inkling, and published the whole thing for anyone to download. Not a demo, not an API: the actual model. This lesson explains what "open weights" means in plain English, why a company would give away its most expensive asset, and where this model will show up in your life even if you never touch it.

01 First, what "weights" are

When people say "the model," they mean a giant file of numbers. Training an AI is the process of tuning billions of these numbers — the weights — until the system gets good at predicting and reasoning. The weights ARE the intelligence: everything the model learned from training is stored in them. Copy the file, and you've copied the model.

Closed labs (OpenAI, Anthropic, Google) guard that file the way Coca-Cola guards its formula. You can talk to ChatGPT all day, but you will never see or change the numbers behind it. Open weights means the finished file is published — anyone can download it, run it on their own hardware, and retrain parts of it for their own purposes. Inkling's weights went up on Hugging Face, a public library for AI models, on day one.

Open weights ≠ open source

You'll see these terms mixed up constantly. Open weights: you get the finished model. Open source (strictly): you'd also get the training data and code — the full recipe, not just the cake. Almost no frontier lab releases the recipe. Inkling is open weights, and that's still the meaningful part for everyone downstream.

02 What Inkling actually is

SpecInklingWhy it matters to you
Size975B parameters (41B "active" per response)It's a Mixture-of-Experts — a committee of specialists where only the relevant few answer each question. Big brain, efficient bills.
InputsText, images, and audioIt can read documents, look at charts, and listen to recordings — one model, three senses.
Context windowUp to 1M tokensRoughly a long novel's worth of material in one conversation.
Thinking effortA dial, not a switchDevelopers can turn "how hard it thinks" up for tough problems or down for speed and cost — per request.
Honesty trainingCalibration & abstentionTrained so that answering only pays off when it's likely right — so it says "I don't know" instead of confabulating.

That last row is the sleeper feature. Most models are trained in ways that quietly reward confident guessing — it's a root cause of hallucinations, which we cover in spotting hallucinations. Thinking Machines trained Inkling with rewards for calibrated answers: confident when it should be, hedged or silent when it shouldn't. A model that knows what it doesn't know is more useful than a slightly smarter one that bluffs.

Thinking Machines' own launch post says it plainly: "Inkling is not the strongest overall model available today, open or closed." Respect for the honesty. Its bet isn't beating GPT or Claude at their own game — it's being the best starting point for organizations that want a model of their own.

03 Why give away the crown jewels?

Because the business isn't the model — it's the customization. Thinking Machines sells Tinker, a platform where organizations fine-tune models on their own data: their documents, their vocabulary, their judgment calls. Inkling is the free foundation designed to pull customers into that workshop (it launched on Tinker with a 50% intro discount, which tells you the strategy). The wider pattern matters for the whole industry: open-weights models mean companies can own their AI — run it on their own servers, tune it to their field, and never worry about a vendor changing prices or shutting an API. That option existed before, mostly from Chinese labs like DeepSeek; Inkling is one of the strongest American entries yet.

04 Where you'll actually meet it

The free playground. Thinking Machines runs an Inkling Playground — a chat interface with web search, free for a limited time — if you want to feel the calibrated-answer style yourself.
Inside other products. Hosting providers (Together AI, Fireworks, Modal, Databricks, Baseten) serve Inkling to app builders. When a niche tool advertises "our own specialized AI," a fine-tuned open-weights model like this is increasingly what's under the hood.
At work. If your employer builds a custom internal assistant trained on company knowledge, open-weights bases like Inkling are exactly what teams reach for — the data never has to leave the building.
Open weights does not mean you can run it at home. At 975 billion parameters, Inkling needs data-center hardware — downloading the file won't help your laptop. A preview of Inkling-Small (a much lighter sibling) is coming, but for individuals the playground and hosted apps are the realistic doors in.

The 60-second takeaway

  • Weights = the file of learned numbers that IS the model; open weights = that file is public.
  • Inkling (July 15, Thinking Machines Lab / Mira Murati) is a big, multimodal, deliberately honest open-weights model — built to be customized, not to top leaderboards.
  • The business logic: give away the model, sell the customization platform (Tinker).
  • For you: expect more products running "their own" AI — and try the free playground if you're curious what calibrated honesty feels like.
Quick answers

Inkling & open weights — FAQ

What is Inkling?
The first AI model built from scratch by Thinking Machines Lab, founded by former OpenAI CTO Mira Murati. Released July 15, 2026 with open weights — the full trained model free to download — it reads text, images, and audio, and was trained to say "I don't know" rather than guess.
What does "open weights" mean?
The weights are the billions of learned numbers that make a model work. Open weights = that finished file is published for anyone to download, run, and modify. Closed models (ChatGPT, Claude) never show you the file; full open source would also include the training data and code — the recipe, not just the cake.
Can I run it on my own computer?
Realistically no — 975B parameters needs data-center hardware. Your doors in: the free Inkling Playground on Tinker (free for a limited time), or apps built on hosted versions from providers like Together AI, Fireworks, Modal, Databricks, and Baseten.
Is it better than ChatGPT or Claude?
No — and the lab says so itself: "not the strongest overall model available today, open or closed." It's built to be the best starting point for customization, not a finished head-to-head assistant.
Up next in AI Foundations

Evaluating new tools: 30 minutes to a verdict

A new model drops every week. Here's the cheap, repeatable filter that tells you adopt, watch, or skip. Read the lesson →