Inkling: Thinking Machines Ships an Open-Weights Model That Fine-Tuned Itself

Thinking Machines, the lab founded by former OpenAI CTO Mira Murati, has released its first model. It is called Inkling, it is fully open weights, and it is available on a public platform right now. The demo the lab led with is the part worth paying attention to: they had the model fine-tune itself, end to end, with no human in the loop.

The video walks through why the lab made an unusual bet. Inkling is not trying to top benchmarks. It is built to be the best model to customize, and the self-fine-tuning demo exists to prove that claim.

Who Built It and Where to Find It

The creator opens with the headline: Mira Murati's new lab just released its first AI model. The company is Thinking Machines, and Inkling is their debut.

The model is fully open weights and lives on a public platform called Tinker, where anyone can experiment with it. The video shows the Tinker playground at tinker.thinkingmachines.ai, with projects, sessions, API keys, and docs visible in the sidebar.

Built for Customization, Not Benchmarks

The lab openly admits Inkling is not the strongest model in the world, and the creator stresses that this is deliberate. Instead of competing on benchmark scores, Thinking Machines built Inkling to be the model you take as a base and train on your own data, your own workflows, and your own product.

The playground settings shown in the video reflect that positioning. The model exposes its system prompt directly, and the interface lists Inkling as a multimodal open-weights model with an April 2026 knowledge cutoff and a 256,000 token context window.

The Self-Fine-Tuning Demo

To prove the customization claim, Thinking Machines had Inkling customize itself. They asked it to become a model that never uses the letter E. The video shows the actual prompt from the lab's demo: "fine tune yourself with Tinker to be a lipogram model where you never use the letter e in your responses, then switch to it."

Inkling then wrote its own training job, ran it, evaluated the results, loaded its new weights, and started answering without a single E. The full loop ran with no human in it.

The Efficiency Angle

The creator flags one more detail for builders: Inkling has a controllable thinking effort dial. In the lab's tests, sweeping that effort setting let Inkling match a leading open model's coding performance while using roughly a third of the tokens, which translates to a third of the cost for the same result.

The creator closes with the strategic read: frontier labs sell you intelligence, while this lab is betting the real market is ownership, meaning models you shape and control yourself.

Key Takeaways

• Thinking Machines, founded by former OpenAI CTO Mira Murati, released its first model, Inkling, as fully open weights on a public platform called Tinker.
• The lab admits Inkling is not the strongest model available. It is deliberately positioned as the best model to customize with your own data and workflows.
• In the flagship demo, Inkling fine-tuned itself into a model that never uses the letter E. It wrote the training job, ran it, evaluated results, and loaded its own new weights with no human involvement.
• Inkling has a controllable thinking effort dial. In the lab's tests it matched a leading open model's coding performance using roughly a third of the tokens, and therefore a third of the cost.
• The creator frames the release as a bet on ownership over raw intelligence: models you shape and control rather than rent.

Resources

• Tinker playground by Thinking Machines shown in the video as the public platform hosting Inkling.