🤚 The Open-Palm Illumination

Former OpenAI CTO Mira Murati has released her first model from Thinking Machines Lab, and it is a monument to competitive spite dressed in an Apache 2.0 license. The model is called Inkling, it has 975 billion parameters, and it is open-weights — making it the largest open model ever released by a Western AI lab.

For context, Murati left OpenAI in September 2024, founded Thinking Machines Lab in February 2025, raised $2 billion from Andreessen Horowitz at a $12 billion valuation, and shipped a competitive frontier model in approximately nine months. OpenAI took five years. Anthropic took three. Mira did it in the time it takes most startups to finalize their logo.

The technical specifications read like a love letter to efficiency engineers:

• 975 billion total parameters in a mixture-of-experts architecture
• Only ~41 billion active parameters per inference call
• 1 million token context window
• Trained on 45 trillion tokens of text, image, audio, and video
• Trained on Nvidia GB300 NVL72 systems

It reasons natively across all four modalities, though outputs are currently limited to text, code, styled artifacts, and structured data. So it can hear your podcast and see your whiteboard, but it will only respond with very well-formatted opinions.

👐 The Two-Handed Reality Check

Here is where the narrative gets interesting, and by interesting we mean strategically brilliant in a way that should make closed-model executives develop a twitch.

Thinking Machines has openly stated that Inkling is “not the strongest overall model available today, open or closed.” This is the corporate equivalent of walking into a boxing ring, handing your opponent a compliment, and then hitting them with your business model.

Because Inkling’s purpose is not to win benchmarks. It is to be customized. The company claims it uses a third as many tokens as Nvidia’s Nemotron 3 Ultra to achieve equivalent coding performance. In a project with Bridgewater Associates — yes, that Bridgewater — a fine-tuned Inkling scored 84.7% on financial reasoning tests, outperforming top proprietary models at roughly one-fourteenth the cost.

The revenue model is elegant: Inkling itself is free. The money comes through Tinker, the company’s fine-tuning and customization platform. You download the weights. You break them. You reshape them into something that knows your industry. And Thinking Machines charges you for the reshaping, not the clay.

This is the open-source razor-and-blades play, except the razor is a 975-billion-parameter neural network and the blades are enterprise contracts.

🌿 The Gentle Awakening

What makes this story genuinely significant — beyond the spectacle of a former CTO returning to the arena with the entire kitchen — is what it says about the economics of frontier AI in mid-2026.

Murati built Inkling from scratch. She did not fork an existing model. She did not license someone else’s weights and add a hat. She trained a nearly-trillion-parameter model from the ground up in nine months, which means the barrier to building frontier AI has dropped from “nation-state budget” to “one good fundraising round.”

The post-training process is worth noting: Thinking Machines used other open-weight models — including Moonshot AI’s Kimi K2.5 — to generate early training data before large-scale reinforcement learning took over. The open-weight ecosystem is now feeding itself, each new model bootstrapping the next in a cycle that closed labs cannot participate in without, well, opening up.

This is what a flywheel looks like when it begins to spin. And it is spinning in the direction of companies that charge for expertise, not access.

👑 The Crown Verdict

Mira Murati has done something that Silicon Valley almost never rewards: she made the generous strategic choice. Open weights. Apache 2.0. Take it, modify it, deploy it, profit from it. The catch is that the generous choice is also the brilliant choice, because every company that downloads Inkling and discovers it needs customization becomes a potential Tinker customer.

In nine months, Thinking Machines Lab went from incorporation to releasing the largest Western open-weights model. The closed-model labs should not be worried because Inkling is better than their models — it probably isn’t, by their preferred benchmarks. They should be worried because Inkling is good enough, it’s free, and the woman who built it knows exactly where their pressure points are.

She was, after all, the one who helped build them.

Inspired by Mira Murati’s 975B Open Model, Ramin Hasani on Post-Transformer AI, and Demis’ AI FINRA | EP #271 by Peter Diamandis.

Your competitive moat is showing. Customize wisely.