Inkling: The First Open Model from Thinking Machines Labs

A New Bet in the AI Market

In a move that signals a departure from what we have come to expect from major AI companies, Thinking Machines Labs — the startup founded by Mira Murati, former Chief Technology Officer of OpenAI — has launched its first internally developed model under the name "Inkling." What sets this model apart is that it is open-weight, meaning developers and organizations can download and modify it directly without restrictions. This approach differs fundamentally from what companies like OpenAI, Anthropic, and Google pursue.

Technical Specifications: Selective Power and Computational Efficiency

Inkling is built on a "mixture of experts" architecture with a total of 975 billion parameters, yet only approximately 41 billion of those parameters are active for any given task. This design is common in large models because it reduces computational cost and increases speed without sacrificing performance. The model was trained on 45 trillion tokens of multimodal data encompassing text, images, audio, and video, although its outputs at this stage are limited to text, code, and structured data.

Among the most notable claims made by the model's creators is that it consumes one-third the number of tokens required by NVIDIA's Nemotron 3 Ultra to reach a comparable level of performance on coding tasks — a clear practical advantage in production environments.

Not the Most Powerful, but the Most Flexible

The company does not claim that Inkling tops global performance benchmarks. In fact, it openly acknowledges in its blog that the model is "not the most comprehensive-performing model available today, whether open or closed." The bet here is not on absolute superiority, but on balanced performance and adaptability. The model allows users to adjust the level of "thinking effort" up or down depending on the situation — moving quickly through routine tasks and going deep when precision is required — and it automatically signals uncertainty rather than guessing.

It is worth noting that the model was not designed as a ready-to-use product, but rather as a starting point that organizations can refine and customize through the company's "Thinker" platform, built specifically for this purpose. This naturally means that responsibility for the safety and suitability of any modifications rests with the deploying organization, not the developer.

The Rise of the Anti-Closed-Model Movement

This launch coincides with a growing wave of criticism directed at the centralized, closed AI model. Satya Nadella, CEO of Microsoft — which has invested billions of dollars in both OpenAI and Anthropic — has warned that organizations relying on closed models pay a double price: direct subscription fees, plus the surrender of their operational knowledge embedded in prompts and fine-tuning to model makers, who then incorporate it into future releases.

In a similar vein, Clément Delangue, CEO of Hugging Face, predicted that frontier models will recede into the role of sandboxes and high-value tasks, while open and proprietary models dominate actual production environments — precisely the direction Thinking Machines Labs is building toward.

A Practical Case Study: Bridgewater

The most compelling argument in favor of this approach came from a joint project with Bridgewater Associates, the world's largest hedge fund. Researchers from both sides took an open-source model and retrained it on Bridgewater's proprietary financial knowledge. The resulting model achieved a score of 84.7% on financial reasoning benchmarks, surpassing closed professional models, at an operational cost of approximately one-seventeenth of theirs — though these results come from an internal evaluation rather than an independent one.

Legitimate Questions About Funding and the Road Ahead

Thinking Machines Labs boasts that it accomplished in nine months what took OpenAI five years and Anthropic three. Yet several legitimate questions cast a shadow over that claim:

• Funding: Reports pointed to a $50 billion funding round that was taking shape last November, only to stall by January. The company has not disclosed its financial position to date.
• Distillation Training: The company acknowledged that it partially relied on outputs from other open models — including Moonshot AI's Kimi K2.5 — to generate some initial training data, though it promises full self-sufficiency in its next model.
• Infrastructure: The company announced a partnership with NVIDIA to deploy one gigawatt of computing capacity, and confirmed that Inkling was trained entirely on GB300 NVL72 systems — without clarifying how these costs are being covered.

Conclusion: Betting on Specialization, Not Dominance

What Thinking Machines Labs is offering is not a claim to having the world's best model, but rather a different philosophy: AI shaped by an organization's own knowledge outperforms off-the-shelf AI built to uniform standards. If this bet proves right, the true competitive advantage lies not in how much you spend, but in making massive spending unnecessary in the first place.

✦ بقلم فريق دروب أيديا