Thinking Machines Debuts Inkling, Its First Open Model Built for Enterprise Customization

Updated July 15, 2026 6:25 pm

Thinking Machines Lab, the artificial intelligence startup founded by former OpenAI chief technology officer Mira Murati, has unveiled Inkling, its first AI model and its first major public test of a strategy built around customizable, open-weight systems. The release matters because it positions the company squarely against the “one model for everyone” approach used by OpenAI, Anthropic and Google, and instead argues that enterprises will get more value by adapting AI to their own data and workflows.

Inkling is not designed as a consumer chatbot meant to dominate the public conversation. It is a large, open-weight model that organizations can download, modify and fine-tune, with Thinking Machines presenting it as a foundation for private deployment rather than a polished end-product. The debut is also the startup’s clearest sign yet that it intends to compete on enterprise flexibility, model efficiency and customization rather than raw benchmark supremacy.

The model arrived Wednesday, more than a year after Thinking Machines began assembling its technical stack largely out of public view. It follows an earlier research preview focused on “interaction models,” a class of systems intended to speak, listen and interrupt more naturally than standard chatbot products. With Inkling, the company is now showing the core technology behind that philosophy and making a direct case that companies should not be forced to rely on fixed, centrally controlled AI systems.

What exactly is Inkling?

Inkling is an open-weight mixture-of-experts model with 975 billion total parameters, though only about 41 billion are activated for a typical task. That architecture is important: by routing each query through a subset of the full model, the system can preserve scale while controlling the computing cost of inference.

According to Thinking Machines, the model was trained on 45 trillion tokens spanning text, images, audio and video. The company says it can reason across those modalities natively, which places it in a more ambitious category than many text-first systems that later gain vision or audio features as add-ons.

The startup is careful not to oversell Inkling as the market’s strongest AI. In its own materials, Thinking Machines says plainly that the model is not the best-performing option among either closed or open alternatives. That framing is deliberate. The company appears to be prioritizing usefulness, adaptability and operational efficiency over a claim that it has already beaten the frontier labs at their own game.

Why the open-weight choice matters

An open-weight release gives developers and companies direct access to the model’s learned parameters, allowing them to inspect, adapt and deploy it without depending on a vendor’s hosted interface. That distinction is central to Thinking Machines’ business thesis, because it creates room for organizations to own more of their AI stack rather than renting access to a closed system.

For enterprise buyers, that can mean more control over data handling, more tailored performance and more freedom to optimize costs. It also changes the economics. Instead of paying recurring access fees for a proprietary chatbot, customers can use the model in their own environments and spend their money on customization, hosting and internal expertise.

Thinking Machines is arguing, in effect, that the value in AI will increasingly come from adaptation rather than from one universal model that everyone uses in the same way.

That argument is becoming more common across the industry, especially among companies that believe the next wave of AI value will come from specific domains such as finance, healthcare, legal work, manufacturing and internal operations rather than from generic public chat.

How does Inkling fit Thinking Machines’ bigger strategy?

Inkling is the company’s first public proof point after a long period of foundation-building. The release shows that Thinking Machines is not trying to be just another chatbot vendor. Instead, it is building a platform around model customization, with Tinker, its model-tuning product, positioned as the main commercial layer.

The startup’s broader view is that enterprises possess deep, proprietary knowledge that general models cannot fully capture. Rather than expecting a central lab to pretrain a perfect system for every industry and use case, Thinking Machines argues that organizations should shape models themselves so that the AI learns from their own practices, terminology and decision-making patterns.

This is a direct challenge to the model used by the biggest AI companies. OpenAI, Anthropic and Google have largely marketed their products as general-purpose assistants first, then layered in more autonomous capabilities over time. Thinking Machines is reversing that order, starting with a customizable base model and then encouraging companies to build specialized systems on top.

The role of Tinker

Tinker is the key to the business model. Thinking Machines is not presenting Inkling as a simple product to be consumed through a subscription. Instead, the company is offering the model as a starting point for organizations that want to fine-tune it, host it and adapt it through Tinker.

That makes the model itself both a technical milestone and a marketing device. The real revenue opportunity, at least in the company’s current framing, comes from the tools and services around the model rather than from metered access to the model alone.

How efficient is Inkling compared with rivals?

Thinking Machines says Inkling can deliver competitive results with fewer tokens than some rival systems. On one internal benchmark, the company claims the model achieved coding performance comparable to Nvidia’s Nemotron 3 Ultra while using about one-third as many tokens.

Token efficiency matters because it affects both cost and latency. A model that reasons with fewer tokens can often respond faster and more cheaply, which is particularly valuable in business settings where response time and deployment cost matter as much as raw accuracy.

At the same time, the company is not claiming the model leads the field on every metric. Instead, it is pitching a balance of capability, calibration and adaptability. The emphasis on calibrated responses, including explicit uncertainty when the model is not confident, suggests that Thinking Machines wants to differentiate Inkling as a system that knows when not to overstate its certainty.

• 975 billion total parameters
• About 41 billion active parameters per task
• 45 trillion training tokens
• Multimodal reasoning across text, audio, images and video

Why are enterprises paying attention?

Enterprises are paying attention because the market is increasingly split between general-purpose AI and specialized production systems. Many companies experiment with flagship models from the biggest labs, but they often discover that the most valuable applications are the ones tuned to their own data, vocabulary and compliance needs.

That insight is shaping a wider shift in enterprise AI strategy. Organizations are beginning to separate exploratory use cases, where a premium proprietary model might make sense, from production deployments, where control and cost are often more important than access to the latest benchmark winner.

Thinking Machines is betting that this split will only deepen. Its view is that a model that can be modified by the customer will often outperform a generic model that has been optimized for broad appeal. The company’s thesis is not that every enterprise needs to become an AI lab, but that serious buyers will want more ownership over the intelligence they depend on.

What Microsoft, Hugging Face and Bridgewater suggest

Recent comments from other industry leaders appear to support that direction. Microsoft chief executive Satya Nadella warned in a weekend blog post that enterprises using proprietary AI are paying twice: once for access and again by effectively donating business knowledge through prompts, corrections and usage patterns that may shape future model improvements.

Hugging Face chief executive Clem Delangue has also argued that frontier models are likely to become tools for experimentation and high-value tasks, while most production systems move to private or open-source alternatives. That is close to the market shape Thinking Machines is trying to build around.

The strongest real-world example so far came from Bridgewater Associates, the world’s largest hedge fund. In a joint project, researchers from Bridgewater and Thinking Machines took an open-source base model and refined it using Bridgewater’s own financial expertise. The companies said the result outperformed leading proprietary systems on financial reasoning tests while running at a fraction of the cost. Those figures were published by the companies themselves and have not been independently verified, but they illustrate the appeal of domain-specific tuning.

What is the company’s speed claim?

Thinking Machines is also highlighting how quickly it reached this stage. The startup says it took about nine months to get from formation to a public model release, a pace it contrasts with the much longer timelines seen at older AI labs.

The comparison is designed to show momentum, but it also reflects a different starting point. OpenAI and Anthropic were building highly visible consumer and developer products while also pursuing frontier research. Thinking Machines appears to have spent much of its early life assembling infrastructure, teams and training pipelines before revealing the model that sits on top of that foundation.

That invisible buildout matters because the company has been relatively quiet about the mechanics of how it got here. Inkling is therefore not just a model launch; it is an assertion that the organization has already achieved enough technical maturity to compete, even if its public footprint remains smaller than that of the best-known AI companies.

How was Inkling trained?

Inkling was pretrained from scratch, according to Thinking Machines, but the company acknowledges a hybrid approach during early post-training. It says it used other open-weight models, including Moonshot AI’s Kimi K2.5, to help generate some of the post-training data before large-scale reinforcement learning took over.

That disclosure is important because the AI sector has been debating how much it should rely on model-generated data, particularly when one system’s outputs are used to improve another. The practice, often described as distillation, has drawn industry scrutiny and raised legal, technical and competitive questions.

Thinking Machines says the next version of its model line will use a fully self-contained post-training process. That suggests the company is aware of the concern and wants to move toward a cleaner pipeline, even if the current model reflects some external assistance.

What questions remain about training data?

The biggest unanswered question is not whether the model is capable of good performance, but how much of that performance comes from outside help, internal engineering and large-scale reinforcement learning. The company has disclosed enough to show it is taking a pragmatic route, but not enough to eliminate all skepticism about how much it borrowed from the broader open-weight ecosystem.

That uncertainty is unlikely to slow enterprise interest, though. For companies deciding whether to test Inkling, the central issue is whether the model can be adapted effectively to their own domain and whether the economics work better than the alternatives.

What does Inkling say about AI economics?

Inkling underscores a broader economic shift in AI: the value may move away from the model itself and toward the services, customization and infrastructure that surround it. If a model is open-weight, anyone can use it, which means the vendor cannot rely on exclusive access as its main source of leverage.

That has major implications for business models. Open-weight AI can compress subscription revenue but increase demand for support, hosting, deployment tooling and fine-tuning services. In that world, the company winning the market is not necessarily the one with the best chatbot interface; it may be the one that makes private AI easier to operationalize.

Thinking Machines appears comfortable with that trade-off. If more enterprises want to own their own models, then the value may shift to platforms like Tinker that help them do exactly that. The company is, in effect, betting that model openness will create a larger ecosystem even if it limits direct monetization of the model release itself.

Can open weights really replace subscriptions?

Open weights can reduce dependence on subscription platforms, but they do not eliminate spending. Organizations still have to pay for hosting, integration, security, internal engineering and ongoing tuning. The difference is that those costs may go to the enterprise’s own infrastructure stack or to vendors like Thinking Machines that help maintain it, rather than to a closed-model provider alone.

That is why the release matters beyond the technical details. It signals that at least some AI startups now believe the market is moving toward ownership and customization, not just access and convenience.

How much money is Thinking Machines spending?

Thinking Machines has not provided a detailed breakdown of its spending, and the company has been cautious about discussing funding. What is known is that it partnered with Nvidia in March to deploy a gigawatt of Vera Rubin computing capacity, and it says Inkling was trained on Nvidia GB300 NVL72 systems.

Those details suggest a serious infrastructure commitment. They also highlight the tension between a company that talks about efficiency and a business that still needs massive computing resources to train a 975-billion-parameter model.

Reports last year suggested the startup was seeking a huge fundraising round, but later accounts indicated that effort had stalled. Since then, the company has not offered fresh public guidance on how much capital it has raised or how it plans to balance spending with revenue growth.

Thinking Machines has signaled that it may not need to spend like the largest closed-model labs if customers are willing to download and run the model themselves.

That is a subtle but significant claim. If true, it could allow the company to avoid the same financial arms race that has defined the frontier AI race so far.

Who is behind Thinking Machines now?

The company’s headcount is now about 200, according to people familiar with the business. That figure is notable because it comes after a period of turnover earlier in the year, including the departure of two co-founders who later joined OpenAI.

Even with those changes, the company is presenting itself as stable and process-driven. One person familiar with the organization said its culture is designed to emphasize continuity rather than personality-driven hype. That approach fits the broader message behind Inkling: the product, not the founder narrative, is supposed to be the story.

That said, Murati’s profile still looms large. As with many AI startups, the founder’s reputation remains a powerful asset, especially in a market where talent, compute and credibility are all scarce.

Why this launch matters for the AI race

Inkling matters because it adds a new and well-funded competitor to the increasingly crowded debate over what enterprise AI should look like. The biggest labs still dominate mindshare with polished assistants and broad consumer reach, but they are no longer the only firms defining what useful AI can be.

Thinking Machines is now making a structural argument: the future is not necessarily a single universal model that everyone accesses through one interface. It may instead be a collection of adaptable systems, owned or controlled by the organizations that use them.

If that thesis proves right, the most valuable AI products may not be the ones that generate the most headlines. They may be the ones that integrate most deeply into a company’s own expertise and workflows, quietly compounding value over time.

For now, Inkling is a first step rather than a final verdict. But it is a highly visible one, and it lands at a moment when the industry is already questioning whether the next phase of AI competition will be about bigger models, better products or more control.

Timeline: how Thinking Machines got here

What happens next?

The next phase will likely be less about the model’s initial release and more about how companies use it. If enterprises begin fine-tuning Inkling for finance, operations, customer support or regulated workflows, Thinking Machines could quickly validate its thesis that open, adaptable models are the real prize.

The company also has to prove that the business works. Open weights can generate adoption quickly, but they can also make monetization harder. The company will need Tinker and its hosting ecosystem to convert technical interest into recurring revenue.

Finally, Thinking Machines will need to show that its future models can improve without leaning on outside assistance. The company’s promise of a more self-contained training pipeline will be watched closely by competitors and customers alike.

For now, Inkling is less a conclusion than a declaration of intent. It says that Thinking Machines wants to redefine what a frontier AI company can sell, who it should serve and how much control customers should have over the systems they rely on.