A Different Kind of AI Lab: Thinking Machines Bets on Customization Over One-Size-Fits-All

When a former OpenAI CTO launches her first AI model, the tech world pays attention. But when that model comes with a disclaimer that it's "not the strongest overall model available today, open or closed," you have to wonder what the strategy really is.

For Mira Murati and Thinking Machines Lab, that admission isn't a failure. It's the whole point.

The Model: Big, But Not in the Way You Think

Inkling is a 975-billion-parameter mixture-of-experts model, but it only activates about 41 billion parameters per task. That MoE architecture borrowed from DeepSeek-V3 keeps it fast and cheap to run. It was trained on 45 trillion tokens of text, images, audio, and video, supports a million-token context window, and accepts multimodal input.

But here's the honesty you don't expect from an AI lab: Thinking Machines explicitly says Inkling isn't the best. It's not going to top the leaderboards. What it offers instead is adaptability . "Inkling is not the strongest overall model available today, closed or open," they wrote in their own announcement. That's not a disclaimer. It's the strategy.

The Bet: Customizable Beats Capable

Most AI labs compete on benchmark scores. Thinking Machines is betting on a different equation: a fine-tuned domain-specific model beats a general-purpose one, even if the general-purpose one scores higher on public leaderboards.

The evidence came from Bridgewater Associates, the world's largest hedge fund. Researchers took an open-source model and trained it further on Bridgewater's own financial expertise. The result scored 84.7% on financial reasoning tests, beating top proprietary AI models, while costing roughly a fourteenth as much to run. Those numbers come from the two companies' own evaluation, but they make the argument clear: for organizations with proprietary data, customization isn't just better. It's cheaper.

This is why Inkling is released under Apache 2.0, the most permissive open license available. Download the weights, fine-tune on your data, deploy in your own infrastructure. No usage restrictions. No per-token API fees. No data leaves your environment.

The Business Model: Not Selling the Model

Revenue doesn't come from Inkling itself. It's free. Revenue comes through Tinker, the fine-tuning platform. The model opens the door. The service is where the business lives. It's a bet that organizations will pay for customization, not for access.

The numbers suggest this could work. One analysis estimated that Inkling uses about a third as many tokens as Nvidia's Nemotron 3 Ultra to hit the same coding performance. Over time, that efficiency advantage becomes a cost advantage that's hard to ignore.

The Cost of Building Fast

Thinking Machines built Inkling in about nine months, much faster than OpenAI's roughly five years or Anthropic's three. To get there, they took a shortcut: they used other open-weight models, including Moonshot AI's Kimi K2.5, to generate early post-training data. The company says it pre-trained Inkling from scratch but used distillation for the initial supervised fine-tuning phase. The next model, they promise, will use fully self-contained post-training.

The hardware is equally ambitious. Thinking Machines partnered with Nvidia for a gigawatt of Vera Rubin computing power and trained Inkling entirely on Nvidia's GB300 NVL72 systems. The BF16 checkpoint requires at least two terabytes of GPU memory roughly eight Nvidia B300s or sixteen H200s. That's not a machine you run in your garage.

The Human Element

After a wave of departures earlier this year, including two co-founders who left for OpenAI in January, Thinking Machines has rebuilt to about 200 employees. Murati's name still carries weight, but the company's culture reportedly favors continuity over reliance on any one personality . It's a remarkable thing for a company founded by someone that famous to insist on.

Where Inkling Actually Stands

On benchmarks, Inkling is competitive. It scored 97.1% on AIME 2026 (mathematical reasoning), 87.2% on GPQA Diamond (graduate-level science), and 77.6% on SWEBench Verified (software engineering). For an open-weights model, those are strong results. But it trails proprietary models from OpenAI and Anthropic on most metrics.

The question is whether that matters. For a developer building an internal tool, the difference between 85% and 90% on a benchmark matters less than knowing the model can be adapted to specific needs. Thinking Machines is betting that the market is ready to trade benchmark leadership for customization.

Microsoft CEO Satya Nadella warned recently that enterprises using proprietary AI models effectively pay twice: once in subscription costs, and again by handing over business knowledge embedded in their prompts . Hugging Face CEO Clem Delangue predicted that frontier models will increasingly be reserved for experimentation while most production AI shifts to private or open-source alternatives .

Inkling is the first real test of that thesis from a major Western lab. It's too early to call it a success. Thinking Machines still needs to prove Tinker can generate real revenue, and Inkling-Small is still in preview. But the model exists, it's free to download, and organizations can fine-tune it on their own data without sending it to an API.

That alone makes Inkling something new in the American AI landscape: a model built for the enterprise, not for the benchmark leaderboard, and open enough to matter.