Mira Murati's Thinking Machines Releases Its First AI Model

30-Second Summary

Thinking Machines Lab released Inkling on 15 July 2026, its first in-house AI model. Inkling is open-weight, meaning companies can download and modify it. It uses a mixture-of-experts architecture with 975 billion total parameters, drawing on roughly 41 billion for any given task. It was trained on 45 trillion tokens spanning text, image, audio, and video. Thinking Machines is positioning it explicitly as a customisation starting point, not the most capable model on the market, and has built a companion platform called Tinker to help organisations adapt it to their specific needs.

At a Glance

• Topic: Model Releases
• Company: Thinking Machines Lab
• Date: 15 July 2026
• Announcement: First public AI model release from the startup founded by former OpenAI CTO Mira Murati
• What Changed: An open-weight, natively multimodal model at frontier scale is now available for organisations to download, fine-tune, and self-host
• Why It Matters: It introduces a credible open-weight alternative to closed frontier models at a scale previously only available through paid APIs, with a companion fine-tuning platform designed for non-expert teams
• Who Should Care: Business operators handling sensitive data, companies in regulated industries, and any organisation that wants AI tuned to its own workflows rather than a general-purpose tool

Key Facts

• Inkling uses a mixture-of-experts architecture: 975 billion total parameters, approximately 41 billion active per inference
• Trained on 45 trillion tokens of text, image, audio, and video natively, not retrofitted with multimodal adapters
• Open-weight: available for download, modification, and self-hosting
• Includes a calibrated uncertainty feature, flagging when it does not know something rather than producing a confident incorrect answer
• Offers a variable "thinking effort" dial, allowing operators to trade reasoning depth for speed and cost
• Thinking Machines explicitly states Inkling is not positioned as the strongest model available, open or closed
• Tinker, the company's model customisation platform, is the designed companion for fine-tuning Inkling on organisation-specific data
• Thinking Machines was founded by Mira Murati, who served as CTO of OpenAI before departing in late 2024

What Happened

Thinking Machines Lab has been one of the most anticipated AI startups since its founding. Mira Murati's departure from OpenAI, where she served as CTO and briefly as interim CEO during the November 2023 board crisis, drew significant attention to whatever the company would build. On 15 July 2026, that question was answered with the release of Inkling.

The model is notable for what it is not. Thinking Machines did not enter the market claiming a top benchmark position or competing directly with GPT-5.6, Claude Sonnet 5, or Gemini 3.5 Pro on leaderboard metrics. Instead, the company positioned Inkling as a starting point, something an organisation modifies rather than consumes off the shelf.

The architecture is a mixture of experts: a large total parameter count (975 billion) that routes each query to a smaller active subset (approximately 41 billion parameters), keeping inference costs manageable at what would otherwise be an extremely large model scale. Training spanned 45 trillion tokens across text, image, audio, and video in native fashion, meaning the model processes all four modalities in a unified way rather than through bolt-on adapters.

Two features stand out as design philosophy signals. The calibrated uncertainty capability is intentional: the model is trained to flag what it does not know, a direct contrast to the confident-but-wrong behaviour that has created liability exposure for organisations deploying frontier models in high-stakes contexts. The variable thinking effort dial reflects a practical understanding of operating costs: not every query requires extended reasoning, and letting operators set that threshold is a lever for managing AI expenditure at scale.

Why It Matters

Open-weight changes the data sovereignty calculation. Most enterprise AI deployments today involve sending queries to a third-party API. That means customer data, internal documents, and strategic information travels to an external server before an answer comes back. Open-weight models eliminate that step entirely. Inkling can run inside an organisation's own infrastructure, under its own security controls, with no external data transfer required.

The companion customisation platform lowers the fine-tuning barrier. Open-weight models have historically required significant machine learning expertise to adapt. Tinker is designed to make that accessible to organisations without dedicated AI research teams. This is the difference between open-weight as a theoretical option and open-weight as a practical tool for a 50-person business.

Native multimodal training at this scale is rare outside closed labs. Most models available for self-hosting are text-first with multimodal capabilities added later. A model trained natively on text, image, audio, and video at 975 billion parameter scale gives operators a genuine foundation for workflows that span document analysis, image interpretation, audio transcription, and video understanding without switching between specialised tools.

The calibrated uncertainty feature matters for regulated industries. Legal, financial, and healthcare operators have faced real problems with AI systems producing confident incorrect output. A model trained to say "I am not certain" is a fundamentally different risk profile in those contexts.

Thinking Machines is signalling a market segment gap. The explicit "not the strongest model" positioning is not a weakness admission. It is a direct appeal to organisations for whom the strongest model available is less important than the most controllable, most adaptable, and most securely deployed model available.

The David and Goliath View

The frontier model race has produced extraordinary capability, but it has also produced a centralisation problem. The organisations building the most capable AI systems are also the ones holding your data, setting your terms of service, and deciding when and how to change the pricing. Inkling is an early and credible bet that a meaningful portion of the enterprise market will eventually decide that control matters more than the last few benchmark points.

For operators running 10 to 200-person businesses, this is not yet a straightforward recommendation. Self-hosting a model at this scale requires infrastructure investment, and fine-tuning requires data preparation and iteration. But the existence of a well-resourced, credibly led startup building deliberately for the customisation market, rather than the benchmark leaderboard, is a signal that the open-weight enterprise segment is becoming commercially viable in a way it was not two years ago.

Watch the Tinker platform closely. The model is the foundation. The tooling is what determines whether organisations outside the Fortune 500 can actually put it to work. If Tinker delivers on accessible fine-tuning, the barrier to operating your own domain-specific AI drops significantly.

Where This Fits in the AI Stack

Inkling enters a market with several distinct tiers. At the top, closed frontier models from Anthropic, OpenAI, and Google offer maximum capability with API-based access. In the middle, large open-weight models like Meta's Llama series and now Inkling offer self-hosting at significant scale. At the base, smaller specialised models handle narrow, high-volume tasks.

Inkling's architectural position is unusual: 975 billion total parameters is a frontier-class scale, but the mixture-of-experts approach keeps active inference cost closer to a mid-size model. That gap has not previously been available in an open-weight format at this parameter count. Combined with native multimodal training, it positions Inkling as a genuine option for organisations that have previously had to choose between capability and control.

Questions Operators Are Asking

Is Inkling production-ready out of the box? Thinking Machines has released it for broad access, but the company's own framing suggests it is designed to be customised before production deployment. Organisations with generic use cases and no sensitive data constraints can use it immediately; organisations with specific domains will get more value from fine-tuning via Tinker.

What infrastructure does self-hosting require? Running a 975-billion parameter model at the full scale requires significant GPU resources. The mixture-of-experts design means active inference only uses roughly 41 billion parameters per query, which reduces the memory requirement substantially, but this is still an enterprise infrastructure conversation, not a single-machine deployment.

How does the data sovereignty benefit hold up under scrutiny? Self-hosting removes the external API risk entirely. Any data processed stays within your infrastructure. This is not a theoretical benefit: it is a material difference in how data governance, client confidentiality obligations, and regulatory compliance work in practice.

Is Tinker available now? Thinking Machines released Inkling with references to the Tinker customisation platform as the companion tool. Operators should check the current availability and pricing of Tinker directly, as release timing and access tiers may differ from the base model.

How does Inkling compare to Meta's Llama models for enterprise self-hosting? Llama is more established with a larger ecosystem of tooling and deployment guides. Inkling enters with a stronger founding team profile and native multimodal training at a larger parameter count. The Future of Life Institute's Summer 2026 AI Safety Index rated Meta at D+, a consideration for operators evaluating open-weight options on governance grounds.

Citable Summary

On 15 July 2026, Thinking Machines Lab released Inkling, the company's first AI model. Inkling is an open-weight mixture-of-experts system with 975 billion total parameters and approximately 41 billion active per inference. It was trained on 45 trillion tokens of text, image, audio, and video. The model is available for download and modification, and is designed for fine-tuning through the company's Tinker platform. Thinking Machines explicitly positions Inkling as a customisation starting point rather than the most capable model in the market. The company was founded by Mira Murati, former CTO of OpenAI.