Inkling and Tinker: Custom AI Training Just Became a Business Decision

Every few months a business owner tells me the same story. They tried AI. It was impressive for a week. Then it wrote a proposal in the wrong voice, quoted a policy the company retired in 2023, and invented a product line that does not exist. The tools got shelved, and the conclusion got written: AI does not really know our business.

That conclusion used to be mostly true, because your realistic options were two. Rent a frontier model that knows everything about the internet and nothing about you. Or start a machine learning project no small business could staff.

This summer, the space between those options got a lot more interesting. Thinking Machines Lab, the AI company founded by former OpenAI CTO Mira Murati, shipped two things that belong together: Inkling, an open-weights model released in July 2026, and Tinker, the training platform built to customize it. Together they turn "train an AI on our business" from a research project into something closer to a procurement decision.

Here is what they are, why the pairing matters, and, because honesty is the product, when fine-tuning is still the wrong tool.

What Inkling actually is

Inkling is Thinking Machines' first model trained from scratch, released under the Apache 2.0 license. You can download it, run it, modify it, and use it commercially without a per-seat contract with anyone.

Three things about it matter for a business:

It is big, but it runs lean. Inkling is a mixture-of-experts model: 975 billion total parameters, but only about 41 billion doing work on any given token. Think of a firm with hundreds of specialists on staff where only the relevant few bill hours on your question. Big-model quality without big-model compute on every request. There is also a lighter preview, Inkling-Small, for work where cost and speed matter more than ceiling.

It reads more than text. Inkling accepts text, images, and audio natively. Scanned invoices, photos of equipment, recorded calls. The messy formats your data actually lives in go straight in, and text comes out.

It has a dial for thinking. You can spend more compute on hard questions and less on routine ones, instead of paying the same price for "summarize this email" and "reconcile these three contracts."

Analysts called it the leading U.S. open-weights model at release. But Thinking Machines was direct that topping leaderboards was not the point. Inkling is a foundation you adapt. That is where Tinker comes in.

What Tinker actually is

Tinker is a training API. You bring your data and your goal. Tinker brings the GPUs, the distributed-training plumbing, and the failure recovery. All the infrastructure that used to require an ML platform team before you could run a single training job.

Under the hood it uses LoRA fine-tuning, which adjusts a small, efficient layer of the model instead of retraining the whole thing. The practical consequences:

• Training is affordable. You are tuning a sliver of the model, not rebuilding a data center's worth of weights.
• You see the process. Tinker exposes the actual training steps instead of an upload-and-pray button, and its open cookbook ships working recipes for the common cases: your document formats, your tone, your workflows, even teaching the model what "good" looks like to your reviewers.
• You keep the result. When training finishes, you download the tuned model and run it wherever you want. Including on hardware you own, inside your own network.

That last point is the difference between renting intelligence and owning an asset.

Why the pairing changes the math

I have written before that most businesses need a capable model inside their own boundary more than a frontier model in someone else's cloud. Inkling plus Tinker takes that one step further. Now the model inside your boundary can be your model.

Your data trains it, and your data stays yours. The path most businesses drift into is pasting sensitive material into a hosted chatbot forever, one prompt at a time. Fine-tuning inverts that. One deliberate, governed use of your data produces an asset you keep, instead of a permanent leak that produces nothing.

The tuned model can come home. Open weights plus an exportable checkpoint means the deployment story ends inside your firewall if you want it to. If you are facing a CMMC assessment, HIPAA obligations, or a bank examiner asking where the AI's data goes, "it runs on our hardware, and here is the training data inventory" is an answerable question. "Ask our vendor" is not.

No permanent vendor dependency. Apache 2.0 means the foundation model cannot be repriced or deprecated out from under your workflow. The tuned model is a file you possess. If Thinking Machines vanished tomorrow, your model would still boot on Monday.

The specialist beats the generalist at your job. A tuned mid-sized model that knows your terminology, your formats, and your rules will routinely outperform a giant generic model at your specific tasks, at a fraction of the running cost. You do not need the model that can do everything. You need the model that does your thing correctly every time.

When fine-tuning is the wrong answer

Now the caveats, because this is where I earn my fee.

If the problem is "the AI doesn't know our current prices or policies," you do not want fine-tuning. You want retrieval: letting the model look things up in your documents at question time. Fine-tuning bakes in style, structure, and skills. It is a terrible way to store facts that change monthly. Most businesses should do retrieval first and fine-tune second, if at all.

If you have no data hygiene, you have a prerequisite, not a project. Training on your data means knowing what is in your data. Customer records, HR files, and privileged material need to be identified and either excluded or approved before a single training run. That governance work is what makes the resulting model defensible.

If nobody will measure it, do not train it. A fine-tune without an evaluation set is a vibes project. Write down twenty to fifty real examples of your tasks with known-good answers before you spend a dollar on training. This is boring. It is also the whole game.

The honest sequence for most organizations: pick one repetitive, high-volume task. Try prompting and retrieval against an open model. Measure. Reach for Tinker only when you hit the ceiling, when the model keeps getting your formats, your tone, or your judgment calls wrong no matter how much context you feed it. That ceiling is real, and plenty of businesses are already standing on it. The news is that there is now a practical next step.

The quiet shift underneath the announcement

Step back from the product names and the shift is bigger than either of them. For years, the AI industry's message to businesses was: the model is ours, the customization is shallow, and the data flows one way, toward us. Thinking Machines built a company around the opposite bet. The durable value is in models adapted to specific businesses, owned by the people they are adapted for.

Whether or not Inkling is the model you end up running, that bet is good news for anyone who would rather own an asset than rent a dependency. The frontier labs still own the ceiling. But taking a genuinely capable open model, teaching it your business, and keeping the result inside your own walls just moved from "research lab" to "line item."

If your team has hit the ceiling of generic AI, the wrong voice, the outdated answers, the made-up product lines, Edwards Consulting Group can help you decide whether retrieval, fine-tuning, or both fit the problem, run the data-governance work that comes first, and stand up the result on infrastructure you control.