Mira Murati, the former CTO of OpenAI, just released her first AI model — and it is built around a very different premise from the products she helped create.

Thinking Machines Lab launched Inkling on July 15, 2026, a 975-billion-parameter multimodal model released under the Apache 2.0 license. The company’s bet is that enterprises are less interested in the most capable general-purpose AI on a leaderboard than in a model they can genuinely make their own.

What Inkling Is

Inkling uses a mixture-of-experts architecture. It has 975 billion total parameters but only draws on 41 billion active parameters during any given request, which keeps compute costs manageable without sacrificing the model’s breadth.

The model was trained on 45 trillion tokens spanning text, images, audio, and video, making it natively multimodal from day one. It supports a context window of up to one million tokens. A companion model, Inkling-Small, packs 12 billion active parameters for teams that need faster, cheaper inference on more constrained workflows.

Both models are open-weight, meaning developers and data teams can download the full model weights and modify them without restriction. Weights are available now; Thinking Machines has committed to releasing open weights for its larger Kimi K3-scale model by July 27.

The Business Model Behind It

The release itself is free. Thinking Machines is not monetizing Inkling through metered API access. Instead, the company generates revenue through its fine-tuning platform, Tinker, where enterprises can customize Inkling on their own proprietary data using LoRA.

On Tinker, teams can work with either a 64,000-token or 256,000-token version of the model. The resulting fine-tuned checkpoints can be downloaded and self-hosted. Critically, Thinking Machines has stated that user data will not be used to train its own models — a meaningful assurance for businesses handling sensitive data.

Customers already using the Tinker platform include Bridgewater Associates, the world’s largest hedge fund.

Why This Matters for Businesses

The AI industry has spent the past two years telling enterprises to pick a foundation model provider and build on top of it. That approach works until you realize your competitive advantage is now shared infrastructure. Anthropic, OpenAI, and Google all serve your competitors too. Your fine-tuned prompts might make you better at using their model — but the model itself belongs to them.

Inkling flips that equation. A business that fine-tunes Inkling on two years of internal customer data, proprietary workflow logic, and domain-specific knowledge ends up with a model no competitor can replicate, because it was trained on data no competitor has.

There are practical trade-offs. Fine-tuning requires data work, MLOps capacity, and some infrastructure investment. Inkling is not designed to top every benchmark — Thinking Machines is explicit that it optimized for adaptability over raw leaderboard performance. Teams that need frontier reasoning on novel tasks may still need access to a state-of-the-art commercial model.

But for businesses with well-defined workflows and proprietary data — which describes most of Enterprise DNA’s clients — the customization argument is genuinely compelling.

The Open-Weight Shift

Inkling is part of a broader pattern. Open-weight models have matured rapidly. Kimi K3 from Moonshot AI (2.8 trillion total parameters) and Meta’s Llama family have demonstrated that open models can reach commercial-model quality. What Thinking Machines is adding is an enterprise-grade fine-tuning service designed to make open weights practically useful for businesses without large ML teams.

This shifts the conversation from “which AI vendor should we choose” to “which base model should we own.” For companies serious about AI as a strategic asset, that is a meaningfully different question.

Thinking Machines’ Positioning

The company’s angle — “customize, don’t just consume” — is also a response to a real concern in enterprise AI buying: the fear of lock-in. When your AI capability lives inside someone else’s API, your vendor can raise prices, change terms, or deprecate your preferred model version. Open-weight fine-tuning eliminates that exposure.

Murati has not positioned Thinking Machines as a challenger to OpenAI or Anthropic on general-purpose capability. The company is competing on a different dimension: control, privacy, and long-term ownership of AI assets.

What This Means for Business

If your team is still treating AI as a subscription to someone else’s intelligence, this release is worth a serious look. The question to ask is: what data does your business have that no one else has? And what would a model trained on that data be worth?

For businesses already using tools like Power BI, Python, or SQL to extract value from their data, the step to owning a fine-tuned AI model is not as large as it sounds. The infrastructure to do it just got substantially more accessible.

Enterprise DNA helps organizations at exactly this intersection — from building the data foundations that make fine-tuning worthwhile to deploying AI agents that run on proprietary models. If you are thinking through what an open-weight AI strategy looks like for your business, a conversation with our advisory team is a good place to start.