Inkling: Thinking Machines' First Open-Weights Model, and How to Use It

Thinking Machines has released Inkling, its first open-weights model, and it lands as one of the more interesting open releases of the year: a natively multimodal model you can download, run, and fine-tune yourself. There are three ways to get hands-on with it, and this piece walks through all three. You can download the weights from Hugging Face, try it instantly in the Playground, or fine-tune it on your own data with the Tinker platform. I'll also cover what it costs, so you can decide between using a hosted version and training your own.

What Inkling is

Inkling is Thinking Machines' first open-weights model, and that's the headline: the weights are published, not locked behind an API. It's natively multimodal, processing text, image, and audio and producing text, and it's built as a large mixture-of-experts (MoE) model, the architecture most frontier labs now use to get more capability per unit of compute. Thinking Machines positions it at frontier scale, and has described it as a trillion-parameter-class model. It ships with a large context window, up to 256,000 tokens, and a knowledge cutoff of April 2026. You can read the full details in the Thinking Machines announcement.

The short version: this is a serious, current, multimodal model you can actually hold in your hands, rather than only rent through someone else's endpoint.

Why an open-weights model matters

An open-weights frontier model is a different proposition from a closed API, and it's the part I care about most given where I work. When you can download the weights, you can run the model on your own infrastructure, keep your data on-premise, and fine-tune it on data that can never leave your walls. For anyone in a regulated or data-sensitive setting, industrial operations, healthcare, anything touched by the EU AI Act's provenance rules, that control is often the deciding factor, not a nice-to-have. Open weights also mean the model can't be deprecated out from under you. What you download, you keep.

Download it from Hugging Face

The weights are published on Hugging Face, which is the fastest route if you want to run or serve Inkling yourself. The Hugging Face release has the model card, the download, and the specifics on license and formats. Check the license there before you build on it. If you'd rather see it work first, start with the Playground below and come back to this.

Try it first in the Playground

Before you download anything, the quickest way to get a feel for Inkling is the Playground. It's a straightforward chat interface with the controls that actually matter: a reasoning-effort slider, from none up to extra-high, so you can trade cost against how hard the model thinks, a max-tokens cap, a web-search toggle, and an editable system prompt. It's the right place to sanity-check whether the model fits your task before you spend anything on compute.

Fine-tune it on your own data with Tinker

The piece that makes Inkling genuinely useful for builders is Tinker, Thinking Machines' API for post-training and fine-tuning. If the Playground is where you try the base model, Tinker is where you make it yours. The workflow is deliberately simple:

• Start from the base model. The Inkling base model is available on the Tinker platform for anyone who wants to customise it.
• Rent the compute. Instead of standing up your own GPU cluster, you rent compute directly from Thinking Machines and train on their infrastructure, which keeps the cost of a fine-tune predictable.
• Bring your data. You point Tinker at your own dataset and go. It's noted as being particularly strong when your data is multimodal, mixing text, image, and audio, which matches how Inkling is built.
• Use the harness. The platform runs on an open code harness for the fine-tuning process, and it goes as far as supporting self-evolution, where a model is directed to fine-tune itself through the Tinker framework.

The public details are light on exact code and UI steps, so treat this as the shape of the workflow rather than a copy-paste tutorial: access the base model, rent compute, bring data, train.

What it costs

Whether you use a hosted version or fine-tune your own, the pricing is worth understanding before you commit. Inkling comes in two context sizes on Tinker, a 64K and a 256K variant, and cost is split across prefill, sampling, and training. Cached prefill gets a steep discount, around 80% off, which matters a lot if you reuse the same long context. At the time of writing there's also a limited-time 50% discount on the model.

The practical read: the 64K variant is the cheaper default, and you only pay for the larger 256K context when a task genuinely needs it, the same route-by-task logic that applies across every model tier, which I wrote about in Fable 5 vs GPT-5.6. If your use case is mostly inference, you're pricing prefill and sampling; if you're customising, training is the line that matters.

Who it's for, and the bottom line

Inkling is aimed at people who want a current, capable model they can own rather than only rent: teams that need to run privately, fine-tune on proprietary data, or build on weights they control. If that's you, the path is clear. Try it in the Playground, download the weights from Hugging Face if it fits, and use Tinker when you're ready to fine-tune. If you only need occasional inference, the hosted pricing may be all you touch.

The bigger picture is that a frontier-scale, multimodal, open-weights model with a friendly fine-tuning path lowers the bar for building something genuinely custom on a strong base. That's the part worth paying attention to: not that another model shipped, but that this one, you can make your own.