A Model That Doesn't Win Benchmarks — And That's Intentional

When Thinking Machines Lab announced its founding last January, the AI world held its breath. Mira Murati, the woman who was at the birth of ChatGPT and led technical development at OpenAI, left the hottest tech company on the planet and took dozens of top engineers with her. Investors poured 2 billion dollars (over 40 billion Czech crowns) into her, and the startup immediately gained a valuation of 12 billion dollars.

But then came silence. A year and a half with no product, no model, just reports of internal disagreements. It wasn't until July 15, 2026, that Thinking Machines finally broke the silence — with the Inkling model. And right from the start, it admitted: it's not the smartest AI on the market.

Inkling is a so-called open-weight model with a mixture of experts architecture. It totals 975 billion parameters — more than GPT-4 — but activates only about 41 billion of them with each query. This means it "wakes up" only the part of the model it needs at that moment, drastically reducing computational costs. In practice, for companies, this means operating costs orders of magnitude lower than with ChatGPT or Claude.

The model handles text, images, and audio and is released under the Apache 2.0 license — one of the most open licenses available. Companies can not only download and operate it for free but also modify and further distribute it, without any licensing fees.

Data as a Weapon — And Why They Bet on It

Why bet on a model that lags behind GPT-5.5 or Claude Opus 4.7 in benchmarks? Murati and her team aren't selling intelligence — they're selling privacy and control.

In recent weeks, a debate has flared up in Silicon Valley about what companies lose when they send their sensitive data to cloud chatbots. Microsoft CEO Satya Nadella speaks of the "reverse information paradox" — according to him, companies pay for AI twice: with money for subscriptions and also with their know-how, which remains in the models. Palantir CEO Alex Karp put it even more bluntly: "Companies are giving away their competitive advantage in exchange for virtually zero value."

And this is precisely the pain point Thinking Machines is targeting. You download Inkling to your own infrastructure — whether you're a bank, a hospital, or an arms manufacturer — and train it exclusively on your internal data. No cloud, no API, no sharing of sensitive information with a third party.

“Human values cannot be averaged. Local knowledge cannot be centralized,” Murati wrote in the manifesto that accompanied Inkling's release. She hit the nail on the head. Venture capitalist Bill Gurley, one of the most respected investors in the industry, commented: “Right place. Right time. Decentralization.”

Bridgewater Tested It — And Saved Thirteenfold

A specific test showed that Murati's and her team's approach really works. According to the Wall Street Journal, engineers at the world's largest hedge fund, Bridgewater Associates, took a freely available model, fine-tuned it using the Tinker platform on their own financial documents, and the result not only beat the most expensive versions of ChatGPT and Claude, but operating costs were approximately thirteen times lower.

Tinker is Thinking Machines' cloud platform, through which companies can securely fine-tune Inkling and other open-weight models on their own data. And this is where the company plans to make money — the model itself is free, but professional fine-tuning and support cost money.

This is a move that makes increasing sense in the European context as well. With the advent of the EU AI Act and the tightening of GDPR, pressure is growing on companies to ensure that sensitive data — especially customer personal data or trade secrets — does not leave European territory. An open-weight model operated on a company's own servers in the EU is exactly what lawyers and compliance departments are calling for.

A Nearly 50 Billion Dollar Valuation That Dissolved

That Inkling even saw the light of day is almost a miracle. When the startup was famously on track last year, it planned another giant investment round that would value it at nearly 50 billion dollars. But this January, that round quietly collapsed. Investors became nervous, and Mark Zuckerberg entered the game with an offer to buy the entire company and its team.

Some of the founders supported the sale, but Murati firmly refused. She wanted independence at all costs — and it almost cost her the company.

As The New York Times revealed, after the disclosure of a secret workplace relationship, Murati limited the powers of key co-founder and CTO Barret Zoph. Zoph couldn't bear the demotion — frustrated by the loss of influence, he secretly arranged a return to OpenAI and, along with two other engineers, gave Murati an ultimatum: either he gets full control over development back, or all three leave.

Murati did not give in to the blackmail. She fired Zoph, the other two left with him, and all three immediately rejoined OpenAI. Thus, the company lost key technical minds within a few days.

What This Means for Czech Companies

Inkling does not yet have Czech localization, and Thinking Machines is not based in Europe, so availability here is not a given. But the principle Murati advocates — AI on your own servers, under your own control — is extremely interesting for Czech companies.

Even today, we have companies here experimenting with open-source models like Meta's Llama or the French startup Mistral. Inkling also falls into this category, and additionally offers a specialized platform for corporate fine-tuning. For a medium-sized Czech company that works with sensitive client data and doesn't want to risk its leakage into American or Chinese clouds, this path is increasingly relevant.

Moreover, with increasing EU pressure on data sovereignty — including the upcoming regulation on European cloud services EUCS — the demand for "download and run yourself" models will only grow. Thinking Machines was one of the first to bet on this trend.