Thinking Machines Lab: Explore Mira Murati’s Vision for Custom AI

Thinking Machines Lab is one of the most closely watched artificial intelligence companies in the world. Founded by former OpenAI technology chief Mira Murati, the company entered the industry with an experienced research team, significant investor backing and an ambitious plan to make advanced AI more useful, adaptable and collaborative.

Unlike companies that focus mainly on building a general chatbot for millions of users, Thinking Machines Lab is developing models and tools that organisations can shape around their own knowledge. Its work combines frontier-model training, model customisation, multimodal interaction and research into how people can collaborate more naturally with AI systems.

The company initially attracted attention because of its high-profile founder and enormous early funding round. It has since moved beyond promises by releasing Tinker, the Inkling model family and a research preview of real-time interaction models. These products reveal a strategy based on giving developers more control over how AI systems are trained and used.

As of August 2026, Thinking Machines Lab remains a young and rapidly changing company. It has powerful financial and infrastructure partners, but it also competes with established laboratories and has experienced several leadership departures. Understanding its products, mission and challenges helps explain why it could become an important force in the next stage of AI development.

What Is Thinking Machines Lab?

Thinking Machines Lab is an artificial intelligence research and product company founded in 2025. Its stated goal is to make advanced AI systems more widely understood, customisable and generally capable. The company believes that people should be able to adapt artificial intelligence to their own needs instead of depending entirely on one fixed model.

Its work covers several connected areas. These include training frontier AI models, providing tools for model fine-tuning, developing multimodal systems and publishing technical research. Rather than treating research and commercial products as separate activities, the company uses product deployment to guide its research priorities and test ideas in realistic environments.

The company has also emphasised human-AI collaboration. It argues that many existing AI systems force users into limited, turn-based conversations in which one person submits a prompt and waits for a completed response. Thinking Machines Lab wants future models to listen, observe, respond and adapt more continuously.

This approach positions the company between a frontier-model laboratory and an AI infrastructure provider. It builds its own models, but it also gives researchers and developers tools for customising models from other organisations. That combination may help it reach customers who want more control than a standard chatbot or closed application can provide.

Who Founded the Company?

Thinking Machines Lab was founded by Mira Murati after she left OpenAI in September 2024. Murati previously served as OpenAI’s chief technology officer and was involved in the development and release of major products, including ChatGPT and GPT-4. Her background immediately gave the new company credibility within the competitive AI sector.

When the company publicly emerged in February 2025, its early team included researchers and engineers who had worked at OpenAI, Meta, Mistral AI and other prominent technology organisations. Reuters reported that approximately two-thirds of its initial team had previously worked at OpenAI, illustrating the strength of its early connection to the ChatGPT developer.

The original cofounding group included experienced figures such as Barret Zoph, John Schulman, Andrew Tulloch, Luke Metz and Lilian Weng. Their backgrounds covered model training, reinforcement learning, safety, infrastructure and AI product development. This concentration of experience helped the startup attract significant attention before it had publicly released a product.

The leadership team has changed since the company’s launch. Several original cofounders and founding employees have left for OpenAI, Meta and other organisations. Mira Murati has remained the company’s chief executive, while PyTorch co-creator Soumith Chintala became its chief technology officer after Barret Zoph’s departure in January 2026.

What Is Its Main Mission?

Thinking Machines Lab says its mission is to build AI that extends human will and judgement. The company believes advanced systems should help people apply their knowledge more effectively rather than replace every human decision with a standardised automated process. This idea influences its products, interfaces and research agenda.

The company argues that valuable knowledge is often local, practical and difficult to place inside a general training dataset. A chef, engineer, doctor or business operator may understand details that are specific to their environment. A general-purpose model cannot automatically capture all this changing, experience-based knowledge.

Thinking Machines Lab therefore wants organisations to customise models using their own data, feedback and goals. Instead of renting the same fixed AI personality and behaviour as everyone else, users could fine-tune systems that reflect their specialised processes. This approach could make models more useful for research, forecasting, search and industry-specific work.

The mission does not reject automation. The company states that machines should perform tasks independently when they can do so reliably. However, it wants models to recognise when human feedback, correction or judgement is required, creating a working relationship that resembles collaboration rather than one-way delegation.

What Is Tinker?

Tinker is the first major product released by Thinking Machines Lab. It is a training application programming interface designed for researchers and developers who want to fine-tune artificial intelligence models. Users control the training process while Tinker manages the computing infrastructure, scheduling and distributed resources behind it.

Fine-tuning allows a developer to take an existing model and train it further for a particular purpose. A company might customise a smaller model for document retrieval, customer support, financial forecasting or tool selection. This can sometimes produce a faster and more affordable system than repeatedly using the largest available general-purpose model.

Tinker provides functions for processing training examples, calculating gradients, updating model weights, generating samples and saving checkpoints. The platform supports supervised fine-tuning and reinforcement-learning workflows. It uses LoRA, an efficient method that trains smaller adapter components instead of updating every parameter in a large base model.

The platform became generally available in December 2025, when Thinking Machines Lab removed its waitlist. The company also added vision-model support and an inference interface compatible with the OpenAI API format, making it easier for developers to connect customised models to existing tools and software.

How Tinker Helps Developers

Training a large AI model requires powerful graphics processors, reliable distributed systems and specialised infrastructure knowledge. These requirements can prevent smaller teams from experimenting with reinforcement learning or large-scale fine-tuning. Tinker attempts to lower this barrier by handling the infrastructure while allowing users to control the training algorithm.

Researchers can use the platform to test new reward functions, distillation techniques or continual-learning methods without building an entire computing cluster. They can choose a supported base model, supply training data or an evaluation environment and then use a relatively small set of functions to control how learning occurs.

Businesses may use Tinker to develop specialised subagents. A search company, for example, could train a smaller model to select documents and tools more efficiently. A forecasting organisation could reward a model according to the accuracy of past predictions, gradually improving its ability to estimate future outcomes from time-stamped information.

Tinker is primarily designed for technically capable users rather than ordinary consumers. Working effectively with it requires an understanding of datasets, model behaviour and evaluation. Its value comes from making advanced training infrastructure more accessible, not from turning model development into a completely effortless or risk-free process.

What Is the Inkling Model?

Inkling is the first major model trained and released by Thinking Machines Lab. The company introduced it on July 15, 2026 as an open-weight, multimodal foundation model. It is designed for general reasoning, coding, tool use, image understanding, audio processing and further customisation through Tinker.

The full Inkling model uses a mixture-of-experts architecture with 975 billion total parameters and 41 billion active parameters during each operation. It supports a context window of up to one million tokens and was pretrained on 45 trillion tokens drawn from text, images, audio and video.

A mixture-of-experts model contains several specialised internal components but activates only a smaller selection for each token. This can provide substantial model capacity without using every parameter for every response. The design aims to balance performance, computational cost and the ability to process long, complicated inputs.

Thinking Machines Lab openly states that Inkling is not the strongest model in every category. Instead, it presents the model as a flexible base with multimodal abilities, adjustable reasoning effort and downloadable weights. This transparent positioning emphasises customisation rather than claiming universal leadership over every closed and open competitor.

What Is Inkling-Small?

Thinking Machines Lab released Inkling-Small on July 30, 2026, only two weeks after the larger model. It is an open-weight mixture-of-experts model with 276 billion total parameters and 12 billion active parameters. The company describes it as approximately one-quarter the size of Inkling while offering comparable results in several areas.

Inkling-Small retains many of the larger model’s important features. It can reason across text, images and audio, supports a context window of up to one million tokens and allows users to adjust how much computational effort it spends on a response. Its smaller active size is intended to reduce latency and operating costs.

According to the company’s evaluations, Inkling-Small matches or exceeds the larger model on several reasoning and agentic coding benchmarks. Inkling still performs better in areas such as broad knowledge coverage and factuality. These results are company-reported and should be considered alongside future independent testing.

The full weights are available for developers, and the model can be fine-tuned through Tinker. This makes Inkling-Small particularly relevant for coding agents, synthetic-data generation, automated evaluation and other workloads in which response speed and cost may matter more than using the largest possible model.

What Does Open-Weight Mean?

An open-weight model makes its trained numerical parameters available for people to download or access. These weights contain the patterns learned during training and allow qualified users to run, study or modify the model. Both Inkling and Inkling-Small have been released with their full weights available.

Open weights do not necessarily mean that every part of a model is fully open source. A company may release weights without providing the complete training dataset, internal infrastructure, data-filtering system or every software component used during development. Users should check the licence and model card before assuming they have unrestricted rights.

For developers, downloadable weights can offer more control over deployment and customisation. An organisation may fine-tune the model around its own tasks, host it within an approved environment or investigate how particular behaviours change after training. This can be attractive when privacy, control or specialised performance matters.

However, running a large open-weight model can still require expensive computing equipment and skilled engineers. Availability does not automatically make deployment easy or affordable. Tinker supports the company’s wider strategy by providing managed infrastructure for users who want customisation without maintaining all the required hardware themselves.

What Are Interaction Models?

Interaction models are Thinking Machines Lab’s proposed approach to more natural human-AI communication. The company introduced a research preview in May 2026. Unlike a traditional chatbot that waits for a complete message, an interaction model can continuously process audio, video and text while responding in real time.

Traditional chat interfaces divide communication into separate turns. The user speaks or types, the model processes the completed input and then generates a response. During generation, the model may not receive new visual or verbal information unless an external system interrupts and restarts the interaction.

Thinking Machines Lab divides incoming and outgoing information into short, time-aligned micro-turns. This allows the model to notice pauses, interruptions, corrections, visual signals and overlapping speech. It can remain involved in a live exchange rather than behaving like an email system that responds only after receiving a completed message.

The company’s system also combines a responsive interaction model with a separate background model for longer reasoning and tool use. The interactive component can continue communicating while the background system searches, plans or performs another complex task. The concept remains a research direction rather than a fully established replacement for existing chat interfaces.

Why Multimodal AI Matters

Multimodal AI can process more than one type of information, such as text, images, audio and video. This matters because people do not communicate only through written prompts. Meaning can also come from tone of voice, visual surroundings, diagrams, facial expressions, timing and physical actions.

Thinking Machines Lab considers multimodality essential for improving human-AI collaboration. A model that can observe a user’s screen, hear spoken corrections and inspect a physical task may receive richer information than one working only from a short text request. This could reduce misunderstandings and make guidance more responsive.

Inkling and Inkling-Small both support native reasoning involving images and audio. The interaction-model research extends this idea by focusing on continuous inputs rather than isolated files. Potential applications could include live translation, collaborative design, technical assistance, education and accessibility tools.

Multimodal systems also introduce serious privacy and safety questions. Continuous audio or video input may contain personal, confidential or irrelevant information. Organisations adopting such systems will need clear consent, secure data handling and controls that determine when the model is allowed to observe, store or act on information.

How the Company Approaches AI Safety

Thinking Machines Lab says it follows an empirical and iterative approach to AI safety. Its public principles include preventing misuse, monitoring systems after deployment, carrying out red-team testing and sharing useful safety techniques with the wider research community.

The company also wants outside researchers to study alignment by providing access to research, code, datasets and model specifications where appropriate. Open research can allow more people to identify weaknesses, test claims and develop improved safety methods rather than leaving all understanding inside a small number of laboratories.

At the same time, open-weight releases create difficult trade-offs. Giving developers more freedom can support beneficial customisation, but downloaded models may be modified in ways the original developer cannot monitor. Strong model capabilities therefore increase the importance of licensing, evaluations and responsible deployment practices.

The company’s broader argument is that alignment should not be controlled entirely by one central model provider. It believes individuals and organisations should be able to shape AI around their own knowledge and values. Whether this decentralised approach can remain both flexible and safe will be one of its most important tests.

Funding and Company Valuation

Thinking Machines Lab raised approximately $2 billion in a funding round announced in July 2025. The round valued the young company at about $12 billion and was led by Andreessen Horowitz. Other participants included Nvidia, Accel, ServiceNow, Cisco, AMD and Jane Street.

The size of the investment was remarkable because the company had not yet released a public product when the funding was announced. Investors were largely backing Murati’s experience, the quality of the team and the potential value of a new frontier AI laboratory.

Large funding is especially important in AI because training advanced models requires substantial spending on chips, data infrastructure, electricity, engineering and safety evaluation. Even billions of dollars can be consumed quickly when a company is attempting to compete with technology organisations that operate enormous computing clusters.

A high valuation also creates pressure. Thinking Machines Lab must demonstrate that its products can attract sustained developer or enterprise demand. Technical achievements alone may not justify investor expectations unless Tinker, Inkling and future systems develop clear markets and reliable commercial models.

The NVIDIA Partnership

In March 2026, Thinking Machines Lab and NVIDIA announced a multiyear strategic partnership. The agreement is intended to deploy at least one gigawatt of next-generation NVIDIA Vera Rubin systems for model training and platforms that provide customisable AI at scale. Deployment is targeted to begin in 2027.

NVIDIA also made an additional investment in Thinking Machines Lab, although the amount was not publicly specified. The two companies plan to work on training and serving systems designed for NVIDIA hardware while expanding access to advanced and open models for researchers, enterprises and scientific institutions.

Access to large amounts of computing capacity can help the company train more capable models and run extensive reinforcement-learning experiments. Inkling itself was trained on NVIDIA GB300 NVL72 systems, demonstrating that the relationship already extends beyond a future hardware commitment.

The agreement strengthens the startup’s ability to compete, but it also highlights how resource-intensive frontier AI has become. A company may have talented researchers and strong ideas, yet still struggle without dependable access to advanced chips, power, networking systems and long-term infrastructure financing.

How It Differs From OpenAI

OpenAI is widely known for consumer and business products such as ChatGPT and its application programming interfaces. Thinking Machines Lab has so far focused more visibly on customisation infrastructure, downloadable model weights and research into continuous human-AI collaboration.

The company’s mission emphasises models that can be shaped by the organisations and people using them. Tinker gives technically capable users control over fine-tuning, while the Inkling releases provide open weights. This differs from relying only on a centrally managed closed model whose deeper behaviour users cannot directly modify.

The two companies are not complete opposites. Both develop multimodal systems, frontier-model capabilities, safety methods and tools for developers. Thinking Machines Lab also added OpenAI API-compatible sampling to Tinker, allowing customised models to connect with software built around a familiar technical format.

Thinking Machines Lab is considerably younger and has a smaller public product portfolio. Its strategy will be tested by whether developers prefer its customisation-focused approach and whether it can keep pace with organisations that already possess major distribution channels, commercial customers and established model families.

Leadership Changes and Talent Competition

The company has experienced substantial changes within its founding team. Andrew Tulloch left for Meta in 2025, while Barret Zoph and Luke Metz departed for OpenAI in January 2026. Zoph’s exit led Mira Murati to appoint Soumith Chintala as the new chief technology officer.

Other founding employees have also joined competing AI organisations. In July 2026, cofounder Lilian Weng stepped down and later returned to OpenAI. Reports indicated that only two members of the original six-person cofounding group remained after her departure.

These departures do not necessarily mean the company cannot succeed. Movement between leading AI laboratories has become common because researchers receive extraordinary compensation offers and opportunities to work with different models, infrastructure and leadership teams. Thinking Machines Lab has continued releasing products despite the turnover.

However, repeated departures can disrupt research continuity, slow projects and create uncertainty for investors or employees. Retaining experienced technical leaders may be as important as raising capital because a frontier AI company depends on specialised knowledge that cannot always be replaced quickly through ordinary hiring.

Main Challenges Facing the Company

The first challenge is competition. Thinking Machines Lab must compete with OpenAI, Google DeepMind, Anthropic, Meta, xAI and well-funded open-model developers. These organisations already operate large research teams, extensive computing infrastructure and products used by millions of people or major enterprises.

The second challenge is converting research into sustainable revenue. Tinker addresses a clear developer need, but the market for model fine-tuning is crowded. Cloud providers, open-source frameworks and model companies already offer training services, managed infrastructure and customisation tools.

The third challenge is proving that open-weight customisation produces enough practical value to justify the additional work. Many organisations prefer a simple hosted model because they do not have the data, engineering skills or evaluation systems required to maintain their own customised version.

The final challenge is balancing openness, safety and control. Thinking Machines Lab wants users to shape models according to their own knowledge and values. It must provide that flexibility without enabling harmful modifications, weak security practices or unreliable models that create serious consequences in high-risk applications.

Why Thinking Machines Lab Matters

Thinking Machines Lab matters because it is testing a different answer to an important question: should advanced AI come mainly from a few centralised systems, or should organisations be able to train and shape models around their own knowledge? Its products place customisation at the centre of the AI experience.

Tinker could make advanced fine-tuning and reinforcement learning available to teams that cannot manage large GPU clusters. Inkling gives developers an open-weight multimodal model designed for further adaptation, while Inkling-Small offers a more efficient option for workloads where speed and operating cost matter.

Its interaction-model research also challenges the standard prompt-and-response interface. If continuous multimodal systems become reliable, users may collaborate with AI through speech, gestures, shared screens and interruptions rather than carefully constructing individual text prompts.

The company is still too young for its long-term impact to be certain. Nevertheless, its combination of model training, open weights, customisation tools and real-time interaction research makes it more than another chatbot startup. Its progress could influence how other laboratories design AI products and developer platforms.

Conclusion: What Thinking Machines Lab Is Building

Thinking Machines Lab is building an ecosystem rather than one isolated AI assistant. Tinker provides training infrastructure, Inkling offers a large open-weight foundation model, Inkling-Small improves efficiency and the company’s interaction research explores more natural ways for humans and machines to work together.

Its central idea is that advanced AI should remain adaptable. Organisations should be able to train models using their own knowledge, modify important behaviours and provide continuous human feedback. This could create systems that are more specialised and relevant than one standard model delivered to every user.

The company has strong advantages, including Mira Murati’s experience, significant funding, capable models and a major NVIDIA infrastructure agreement. It also faces serious challenges involving intense competition, expensive model development, commercial uncertainty and the loss of several founding team members.

Thinking Machines Lab has already moved from secrecy to tangible products, but its larger vision remains unfinished. Its future will depend on whether developers and businesses find lasting value in customisable AI and whether the company can turn impressive research into dependable, safe and widely adopted technology.

Frequently Asked Questions

What does Thinking Machines Lab do?

Thinking Machines Lab develops artificial intelligence models, research and training tools. Its main products include the Tinker fine-tuning platform and the open-weight Inkling model family.

Who owns Thinking Machines Lab?

Thinking Machines Lab is a privately held company led by cofounder and CEO Mira Murati. Its investors include Andreessen Horowitz, NVIDIA, Accel, AMD, Cisco, ServiceNow and Jane Street.

What is Tinker AI?

Tinker is an API that allows researchers and developers to fine-tune large AI models while Thinking Machines Lab manages the underlying computing infrastructure, scheduling and distributed training resources.

Is Inkling an open-source model?

Inkling is described as an open-weight model because its full trained weights are available. Open weight does not always mean every dataset, training process and software component is completely open source.

Is Thinking Machines Lab connected to OpenAI?

It is an independent company, not a division of OpenAI. However, founder Mira Murati and many early team members previously worked at OpenAI, and several founding employees later returned there.