Thinking Machines just dropped Inkling — a 975-billion-parameter open-weights model. The Ink Machine is how you run it on your own key, inside your Agent OS, and feed it one prompt to ink out a finished app or game.
Everything here comes from Thinking Machines' own launch and docs — nothing second-hand. Here's where to read it and run it:
"Inkling is a Mixture-of-Experts transformer with 975B total parameters, 41B active. It supports a context window of up to 1M tokens… we are releasing a model we trained from scratch with the full weights available, so that people can make it their own."
— Thinking Machines Lab, "Introducing Inkling", 15 July 2026
Think about the AI models you use every day.
You don't own any of them.
You rent them, one prompt at a time, from someone else's cloud.
The weights are locked in a black box.
You can't see inside. You can't run it offline. You can't shape it to your work.
The price is whatever they decide, and it changes when they feel like it.
And the moment they retire a model or move it behind a bigger paywall, your workflow breaks.
You built your business on a brain you don't hold the keys to.
The Ink Machine breaks that cycle for good.
It's one line of config and your own key. No servers, no GPUs, no fine-tuning needed.
A full one-shot build cost me pennies — and it runs right inside a tab you already have.
Inkling is a frontier AI model from Thinking Machines — the lab Mira Murati (OpenAI's former CTO) started.
Two things make it different from the models you rent:
1. It's open-weights. The full model is public and free on Hugging Face. Nobody can take it away or hide it behind a paywall.
2. You run it on your own key. You grab a key from Tinker — Thinking Machines' platform — and point any tool at it. The model answers to you, not to a middle-man.
The Ink Machine is the simple move: wire Inkling into the coding tab in your Agent OS, and feed it one prompt. It inks out a finished, working thing — an app, a game, a page — in one shot.
Here's what actually happens on your machine, from key to build. This is the real setup I ran.
Sign in at tinker.thinkingmachines.ai, hit API Keys, create one. It shows the secret once — copy it. This is the only paid part, and it's usage-based (pennies per build).
Inkling is served through an OpenAI-compatible endpoint. So any tool that speaks OpenAI can point at it — no special SDK. One line: the base URL, your key, and the model name thinkingmachines/Inkling.
opencode is the open-source coding tab in the Agent OS. Add Inkling as a provider, and it shows up in the model dropdown right next to the free models.
"Build me a neon starfield with a glowing title." Press Enter. opencode sends it to Inkling on your key.
It reasons through the design, then uses the Write tool to drop a complete, self-contained index.html into your workspace. No copy-paste. It builds the real file.
The build lands in the Workspace tab with a live preview, and every run is saved and logged to your Obsidian vault. Nothing gets lost.
So when someone asks "but what IS it?" — the Ink Machine is a frontier model you own the weights to, running on your own key, printing finished builds inside a tab you already have.
What you're looking at: the opencode tab inside my Agent OS, with Inkling picked in the model dropdown — the sub-line reads tinker/thinkingmachines/Inkling · Tinker · 975B MoE. It sits right next to the free models, in a tab I already had. That's the whole "wiring" — one provider, one dropdown entry.
This is the repeatable play for turning any open-weights frontier model into a build machine you own. Four moves, in order.
Pick an open-weights model you can actually hold — Inkling's full weights are public. You're not renting a black box; you're taking a real, ownable model.
Get your own API key and point the tool at it. The model answers to you — your usage, your key, your control. No per-seat middle-man deciding your price.
Plug it into the coding tab in your Agent OS as one provider. Now the frontier model lives next to your free models, in a workflow you already know.
Feed it one prompt. It reasons, writes the real files, and drops a finished build in your workspace — saved, previewed, logged. One sentence in, a working thing out.
No. I nearly upgraded to Databricks Premium for this — then didn't need it.
Inkling runs straight off a Tinker key on an OpenAI-compatible endpoint. One config line. No enterprise tier, no cloud contract.
I gave Inkling one sentence: a glowing aurora over a starfield with an INKLING title. No edits, no second pass. This is the actual file it wrote, rendering live in my Agent OS.
What you're looking at: a live, animated generative-art page — Inkling one-shot it in 35.9 seconds for about 1,700 tokens (pennies). Complete, valid HTML with a canvas animation, saved straight into the opencode Workspace tab. One sentence in, this out.
Thinking Machines is honest about it: Inkling isn't the single strongest model on Earth. The closed frontier (Fable 5) still tops the charts. What makes Inkling special is the combination — frontier-class, multimodal, and yours to own and run.
No — that's the biggest myth about it. The Agent OS runs the everyday 90% on a free local model on your own machine (nothing leaving it), free APIs slot in for more, and for frontier work it drives the CLIs you already pay for — your Claude subscription already includes the Claude CLI, and the OS plugs straight into it, so you're not paying twice. Inkling here is the same idea: your own key, pennies per build. Inside the Boardroom there are full token-efficiency tutorials so you cut usage to the bone.
You can wire this together yourself with the steps above. Or get the whole thing done inside the Agent Operating System — Inkling, opencode, and your free local models, pre-connected.
Wrong: "Open-weights models are weak toys — the real power is all closed."
Right: Inkling is a 975B multimodal frontier model that beats Nemotron 3 Ultra on agentic coding and sits right next to GPT-5.6 Sol. Open no longer means weak.
Wrong: "To run a frontier model myself I need GPUs, a server, and a PhD."
Right: You need one API key and one line of config. It's served on an OpenAI-compatible endpoint — any tool that speaks OpenAI can use it, no infrastructure.
Wrong: "I'll wait until this stuff is easier and cheaper."
Right: It's already pennies per build and it's live today. The people wiring their own models in now are the ones who'll own their stack when everyone else is still renting.
Members post their wins every day — agency owners, ecom founders, course creators, solo operators across 38 countries. Real businesses, real numbers, in their own words.
Read the 158-page wins doc →You stopped renting. Inkling's full weights are open — a frontier model you actually own, not a black box.
You hold the key. One Tinker key, one config line — the model answers to you, not a middle-man.
You skipped the enterprise bill. No Databricks Premium, no cloud contract — just an OpenAI-compatible endpoint.
You build in one shot. One prompt → a finished, working app or game in ~36 seconds, saved and previewed.
You pay pennies. A full build ran ~1,700 tokens — cents, on your own usage.
It lives in your OS. Inkling sits in the opencode tab next to your free models — one workflow, everything logged.
Inkling gives you a frontier model you own. The Boardroom gives you the whole operating system I run a seven-figure business on — with Inkling, opencode, and every model you already pay for wired into one place.
Readers bookmark this and keep renting. Operators join, wire the Ink Machine in this week, and start printing builds on a model that's theirs.
