- Thinking Machines released Inkling, an open-weight mixture-of-experts model with 975 billion total parameters and about 41 billion active, trained on 45 trillion tokens.
- Inkling is not the strongest model available, and it is not the product. The company earns revenue from Tinker, its fine-tuning tool already used by Bridgewater.
- Founded by former OpenAI CTO Mira Murati, the startup is valued at $12 billion roughly a year after launch.
Inkling ships as an open-weight model built to be changed
On Wednesday, Thinking Machines released Inkling, its first public model, more than a year after Mira Murati left the chief technology officer seat at OpenAI to start the company. Inkling is a mixture-of-experts system with 975 billion total parameters, of which about 41 billion activate for any given query, trained on 45 trillion tokens of text, image, audio, and video. It is open-weight, so developers can download it and customize it without access to the training data or source code.
The specifications are strong without being frontier-leading. Inkling leads the current crop of open models from US labs on several public tests but trails the largest Chinese systems, and it does not match the strongest closed models from Anthropic, OpenAI, or Google. Under the old scoreboard logic, that would read as a miss. It is not, because Thinking Machines is not playing that game.
Why Thinking Machines is not trying to win the frontier
Every frontier lab sells the same promise. Rent our single best model through an API, and trust that it is good enough for your task. Thinking Machines is betting the opposite, that a large share of serious enterprise buyers do not want to rent one general model. They want to own and shape a capable base model around their own data, their own domain, and their own cost ceiling.
Inkling is engineered for that. The model card emphasizes balancing cost against performance and processing queries across media, the features a company cares about when it plans to fine-tune and deploy, not when it wants a chatbot. The open weights are the point. A model you can download is a model you can specialize, run on your own compute, and keep behind your own firewall.
Give away the engine, sell the tuning. Thinking Machines is betting that customization, not raw capability alone, is what enterprises will pay for.
That is the thesis behind releasing a model the company openly admits is not the most powerful one available.
The business model is Tinker, not the model
Here is the part most coverage underplays. Thinking Machines is not trying to monetize Inkling at all. Its revenue comes from Tinker, a developer tool for fine-tuning and customizing models, which it already sells to customers including the hedge fund Bridgewater Associates, which uses it to sharpen AI performance on financial tasks.
That reframes the release. Inkling is not a product line, it is a demand generator. A free, capable, open-weight base model is the on-ramp. The recurring revenue is in the customization layer that sits on top. It is a razor-and-blades structure applied to foundation models, and it aligns neatly with where enterprise AI spending is actually moving, toward specialized deployments rather than general chat.
| Metric | Detail |
|---|---|
| Total parameters | 975 billion, about 41 billion active per query |
| Training data | 45 trillion tokens across text, image, audio, and video |
| License | Open-weight, downloadable and customizable |
| Company valuation | $12 billion |
| Revenue product | Tinker fine-tuning tool, customers include Bridgewater |
What a customization-first lab means for the frontier oligopoly
If the bet is right, it pressures the closed labs on their most comfortable assumption, that owning the single best model is the same as owning the market. A $12 billion company founded by one of the most credible technical leaders in the field is arguing that the value is migrating from the model to the layer that adapts it. Fireworks, which raised $1.5 billion the same week on the back of specialized-model inference, is making a version of the same argument from the infrastructure side.
None of this requires Inkling to be the best model in the world. It requires enough enterprises to conclude that a very good model they can shape beats a slightly better one they can only rent. Thinking Machines has now put a free, capable, open-weight model in the market to prove the point, and priced its actual business one layer up. The frontier labs still have the strongest models. Murati is betting that is no longer the same thing as having the customers.
Santage is committed to independent, transparent journalism. This article is produced in accordance with Santage's Editorial Standards and aims to provide accurate and timely information. Readers are encouraged to verify information independently.