U.S.-based Thinking Machines Lab has released Inkling, a customizable open-weights artificial intelligence model capable of processing text, images, and audio, offering businesses and researchers an alternative to closed AI systems.
Editorial Note
This article discusses a newly released artificial-intelligence model developed by the U.S.-based company Thinking Machines Lab.
Inkling was officially released on July 15, 2026, in the United States. The announcement entered the July 16 news cycle in Japan because of the time difference. This article does not claim that the official U.S. release occurred on July 16.
Inkling is an artificial-intelligence system rather than a human mind, verified source of truth, or replacement for professional judgment. Like other large models, it may produce inaccurate, biased, incomplete, or misleading outputs.
New To Education does not endorse Thinking Machines Lab, its executives, investors, commercial partners, or the use of any specific AI model. Organizations should evaluate security, privacy, intellectual-property, cost, and reliability concerns before deploying artificial intelligence.
The artificial-intelligence industry has spent years competing to build increasingly powerful general-purpose systems.
A new U.S.-developed model is making a different argument.
Thinking Machines Lab, the artificial-intelligence company founded by former OpenAI chief technology officer Mira Murati, has released its first major foundation model, called Inkling.
Inkling can process text, images, and audio and generate written responses. It contains approximately 975 billion total parameters, although only about 41 billion are activated during a particular task through a design known as a mixture-of-experts architecture.
The model is not being presented as the most powerful AI system available.
Thinking Machines Lab instead describes Inkling as a flexible foundation that developers, researchers, and businesses can customize for their own specialized needs.
That strategy challenges a central assumption of the current AI race: that every organization needs to rely on one enormous, centrally controlled model built and operated by a small number of technology companies.
Inkling points toward a different future—one in which organizations can download, adapt, host, and fine-tune artificial-intelligence systems rather than simply renting access to them.
What Thinking Machines Lab Released
Inkling is the first model in a planned family of artificial-intelligence systems from Thinking Machines Lab.
The company describes it as an open-weights, multimodal foundation model.
Open weights means the numerical parameters learned during training are made available so outside developers can run or modify the model under its licensing terms.
That is different from a fully open-source system in which every part of the training process, dataset, code, and infrastructure may be publicly available.
Inkling accepts text, image, and audio inputs and produces text outputs.
A user could potentially provide a document, photograph, diagram, spoken recording, or written question and ask the model to analyze the material.
The company says the model was designed to reason across different forms of information rather than treating every input as isolated data.
Inkling is also available for fine-tuning through Thinking Machines Lab’s Tinker platform, which allows organizations to adapt AI models using their own specialized examples and datasets.
Inkling Is Large but Does Not Use Every Parameter at Once
Inkling contains approximately 975 billion total parameters.
Parameters are the internal numerical values an AI model adjusts during training. They help determine how the system recognizes patterns and generates responses.
A larger parameter count does not automatically mean a model is smarter, safer, or more useful.
Inkling uses a mixture-of-experts architecture.
Under this approach, only certain parts of the model are activated for a particular request. Inkling reportedly uses approximately 41 billion active parameters during each forward pass rather than using all 975 billion simultaneously.
The design can reduce the computing cost required for each response while preserving a large overall capacity.
A simplified comparison would be a large organization with hundreds of specialists.
Not every specialist works on every assignment. The system directs a task toward the internal components considered most relevant.
This can improve efficiency, although it still requires powerful computing infrastructure.
Inkling is far too large for most people to run on a typical laptop or home computer.
The Model Can Work Across Text, Images, and Audio
One of Inkling’s defining features is its multimodal design.
Many earlier AI systems were trained primarily to understand and generate text.
Newer systems increasingly work across several types of information.
Inkling can accept written language, visual material, and audio.
A developer could potentially adapt the model to analyze diagrams, summarize recordings, interpret photographs, answer questions about documents, or combine information from multiple sources.
The model does not currently produce images, speech, or video as its primary output. It generates text.
That still creates a wide range of potential applications.
A company could fine-tune it to review technical documents and corresponding diagrams.
A university research team could adapt it to work with recorded interviews and written transcripts.
A manufacturer could potentially connect images of equipment with maintenance instructions.
An education organization could explore tools that analyze spoken explanations, student work, and instructional materials together.
Those uses would require careful testing and appropriate privacy protections.
Thinking Machines Is Not Claiming Inkling Is the Best AI Model
Technology launches often rely on sweeping claims.
Thinking Machines Lab has taken a more measured approach.
The company openly states that Inkling is not the strongest overall model available, whether compared with open or closed competitors.
Instead, it argues that Inkling offers a useful combination of multimodal capabilities, efficient reasoning, controllable thinking effort, and customizability.
That positioning matters.
The AI industry frequently treats benchmark leadership as the primary measure of progress.
A model may receive attention because it scores slightly higher than a competitor on mathematics, coding, reasoning, or language tests.
Businesses do not always need the system with the highest general benchmark score.
They may need a model that can be adapted to a narrow task, deployed under their own controls, or operated at a manageable cost.
Inkling’s value will depend less on whether it wins every benchmark and more on whether organizations can make it useful for real work.
Customizable AI Is the Central Idea
Thinking Machines Lab is betting that organizations will want AI systems shaped around their own data, terminology, workflows, and standards.
A hospital may need a model that understands its documentation structure and clinical procedures.
A manufacturer may need one familiar with specific machinery, safety rules, and maintenance records.
A financial organization may need a system trained to recognize relevant information while following strict compliance requirements.
A school system may need an AI assistant aligned with its curriculum, accessibility standards, student-protection rules, and local policies.
A general chatbot may understand broad topics while still performing poorly in specialized environments.
Fine-tuning attempts to close that gap.
Organizations provide carefully selected examples showing the model how to handle their particular tasks.
The model’s behavior can then become more consistent with those examples.
Customization is not automatic proof of accuracy.
Poor training data can produce poor results. Biased examples can reinforce discrimination. Confidential information can create privacy risks.
The quality of a customized system depends heavily on the quality of the data, evaluation, and human oversight surrounding it.
Open Weights Give Developers More Control
Closed AI systems typically operate through a company-controlled website or application programming interface.
Users send their requests to the provider’s infrastructure and receive a response.
They generally cannot inspect or directly modify the model’s underlying parameters.
Open-weight models offer greater control.
A qualified organization may be able to host the model on its own infrastructure, modify it, fine-tune it, or connect it with internal systems.
This can support privacy, customization, research, and long-term independence from a single vendor.
It can also create new responsibilities.
Organizations operating an open-weight model must manage security, computing infrastructure, updates, monitoring, and abuse prevention.
Access to the weights does not make the technology simple.
It transfers more control—and more risk—to the organization deploying it.
The Release Challenges Centralized AI
Much of today’s most capable artificial intelligence is controlled by a small number of companies.
Users access these systems under commercial terms that may change.
Providers can adjust pricing, remove models, alter safety rules, restrict features, or discontinue services.
Organizations building heavily around one provider may become dependent on that company’s decisions.
Open-weight systems offer an alternative.
A business may be able to keep operating a downloaded model even when a provider changes its commercial platform.
Researchers may be able to study model behavior more closely.
Developers can build specialized services without sending every request through one central company.
This does not eliminate corporate influence.
Training a model as large as Inkling still requires enormous financial and computing resources.
The hardware needed to operate it at scale remains concentrated among wealthy companies and institutions.
Open weights make the model more accessible, but they do not make advanced AI equally available to everyone.
Inkling Was Developed in the United States
Thinking Machines Lab is a U.S.-based artificial-intelligence company founded by Mira Murati and other researchers with experience at major technology laboratories.
Murati previously served as chief technology officer of OpenAI.
The company attracted significant attention before releasing a major foundation model because it raised approximately $2 billion in seed funding and reached a reported valuation of about $12 billion in 2025.
That level of financing created unusually high expectations.
Inkling is the first major test of whether the company can translate its funding, talent, and public profile into competitive technology.
The model also enters a global AI environment in which American companies face strong competition from laboratories in China, Europe, and other regions.
Thinking Machines reportedly used architectural ideas and training techniques influenced by existing open models, including systems developed outside the United States.
Modern technology rarely emerges in complete isolation.
AI development frequently builds upon research papers, open-source tools, public datasets, previous model designs, and international collaboration.
Inkling is a U.S.-developed model operating within a global research ecosystem.
The Model Was Trained on Multiple Types of Data
Inkling was trained on large quantities of text, images, audio, and video data.
The system accepts text, image, and audio input, even though its output is currently text.
Training across several types of data may help the model develop connections that a text-only system would miss.
For example, understanding language about music, machinery, facial expressions, or physical environments may improve when the model has encountered corresponding audio or visual information.
The company reports that Inkling was pretrained on approximately 45 trillion tokens across its data mixture.
A token is a unit of information used during AI training. In text, it may represent a word, part of a word, punctuation, or another linguistic element.
The enormous scale demonstrates why training frontier AI models requires powerful computing clusters, large energy supplies, sophisticated data systems, and substantial funding.
It also raises questions about where the data came from, whether creators consented to its use, and how copyrighted or personal material was handled.
Copyright Questions Remain Important
Open-weight AI does not escape the broader copyright debate surrounding generative models.
Developers train large systems using enormous collections of digital material.
Writers, artists, musicians, publishers, software developers, and media companies have questioned whether their work was used without permission or compensation.
Thinking Machines provides information about Inkling’s capabilities and intended uses, but outside observers may still want greater detail about the training data.
AI companies often avoid publishing a complete list because datasets are enormous, commercially sensitive, or difficult to document fully.
That lack of transparency creates tension.
Developers want enough information to understand a model’s risks.
Creators want to know whether their work was used.
Companies want to protect their methods and data pipelines.
Courts and lawmakers are still determining how existing copyright law applies to large-scale AI training.
Organizations using Inkling should not assume that open weights eliminate intellectual-property concerns.
Safety Does Not End When a Model Is Released
An open-weight system can be adapted for beneficial purposes.
It can also be modified to remove safeguards or support harmful activity.
Potential risks include fraud, impersonation, automated cyberattacks, misinformation, harassment, privacy violations, and the creation of dangerous instructions.
Thinking Machines Lab published a model card describing intended uses, limitations, and safety considerations.
Model cards help users understand how a system was designed and tested.
They are not guarantees.
Once model weights are released, the original developer may have limited control over how every copy is modified or deployed.
This is one of the central tensions in open AI.
Greater access can support innovation and research.
Greater access can also give malicious actors more freedom.
The policy challenge is finding ways to preserve legitimate experimentation without ignoring foreseeable harm.
The One-Million-Token Context Window Is Significant
Inkling’s downloadable model reportedly supports a context window of up to one million tokens.
A context window describes how much information a model can consider during a single interaction.
A larger context window can allow the system to examine long documents, extensive codebases, multiple reports, or large collections of connected material.
That does not mean the model will remember or interpret every detail perfectly.
Models can lose track of information, overlook contradictions, or emphasize the wrong passages even when the material technically fits within the context limit.
A large window is capacity, not comprehension.
For developers, however, the ability to provide more material at once can be useful.
A model could potentially review a long technical manual, analyze a substantial software repository, or compare multiple research documents without requiring the user to divide everything into very small sections.
Controllable Thinking Effort Could Reduce Costs
Thinking Machines says Inkling allows users to adjust how much reasoning effort the model applies.
Some questions require little analysis.
Others involve multiple steps, complex instructions, or difficult comparisons.
Using maximum computational effort for every simple request would be wasteful.
A controllable reasoning system can potentially respond quickly to routine tasks while spending more time and computing power on difficult ones.
This matters because AI costs are strongly connected to computation.
More reasoning often means greater latency, energy use, and expense.
Organizations deploying AI at scale may need to process millions of requests.
Small efficiency improvements can create substantial financial and environmental differences.
The central question will be whether the model can reliably recognize how much effort a task requires or whether developers must set those limits manually.
Inkling Could Support Robotics and Physical Systems
Multimodal AI may eventually become particularly valuable in robotics.
A robot operating in the physical world must process more than written instructions.
It may need to understand visual scenes, spoken commands, object locations, environmental sounds, and changing conditions.
Inkling is not itself a robot.
Its ability to work with images, audio, and language could support future systems that connect AI reasoning with sensors and machines.
A warehouse robot might interpret spoken directions and visual information about inventory.
An assistive device might analyze an environment and provide written or spoken guidance.
A maintenance system might combine photographs, machine sounds, and technical manuals to identify possible equipment problems.
Physical-world deployment carries higher risks than an ordinary chatbot.
An incorrect text answer may inconvenience a user.
An incorrect robotic action may damage property or injure someone.
Any connection between foundation models and physical systems requires strong safety controls, testing, and human override mechanisms.
Education Could Benefit, but Schools Should Move Carefully
A customizable multimodal model could have educational applications.
A school or university might adapt a system to its curriculum, institutional policies, or specialized subjects.
Students could potentially upload diagrams, spoken explanations, and written work for feedback.
Language learners might practice through audio while receiving written corrections.
Students with disabilities could benefit from tools that connect visual, auditory, and textual information.
Researchers could adapt models to analyze interviews, academic literature, or laboratory records.
These possibilities should not be confused with proven educational outcomes.
An AI tool may generate impressive responses without improving learning.
Students may become dependent on it, accept inaccurate explanations, or use it to avoid completing their own work.
Schools must also protect student data.
Uploading identifiable records, disability information, grades, voices, or photographs into an external AI system may create serious privacy concerns.
Education leaders should require evidence, clear policies, and human supervision before adopting new models.
Smaller Organizations Could Still Be Left Behind
Open-weight technology is often presented as more democratic.
Inkling remains an extremely large system.
Running it efficiently may require clusters of expensive graphics-processing units, specialized engineering, substantial electricity, and advanced security controls.
Large corporations and research universities may be able to build that infrastructure.
Small businesses, schools, nonprofits, and local governments may not.
They may still need to access Inkling through commercial hosting providers or Thinking Machines’ Tinker platform.
This means open weights do not automatically create equal access.
A model may be technically available while remaining practically unaffordable.
Smaller versions, shared public infrastructure, educational grants, cloud credits, and efficient fine-tuning methods may help reduce that divide.
Thinking Machines has previewed a smaller Inkling model, but smaller does not necessarily mean lightweight enough for ordinary consumer hardware.
AI Competition Is Shifting Toward Customization
The first stage of the generative-AI boom focused heavily on general chatbots.
Companies competed to build systems that could answer almost any question, generate writing, create code, and perform broad reasoning tasks.
The next stage may be more specialized.
Organizations increasingly want models that understand their own work.
A legal firm may need different behavior from a hospital.
A university may need different safeguards from a marketing agency.
A manufacturer may prioritize technical accuracy while a customer-service department prioritizes tone and response speed.
Inkling’s release reflects that transition.
The future AI market may not be dominated by one universal model.
It may involve a smaller number of broad foundation models adapted into thousands of specialized systems.
That could create more useful tools.
It could also make oversight harder because every customized version may behave differently.
Businesses Should Not Customize AI Without Governance
Fine-tuning a model can create the impression that an organization has built an expert system.
A model trained on company documents may use the correct terminology while still making serious mistakes.
Organizations need evaluation standards before deployment.
They should test accuracy, bias, privacy, security, reliability, and failure behavior.
They should also determine when human approval is required.
A model helping an employee search internal records presents a different level of risk from one making credit, hiring, medical, educational, or legal recommendations.
Businesses should document which data were used for training, who approved them, and how outputs are monitored.
Customization should not become a shortcut around accountability.
A company remains responsible for the system it deploys, even when the base model came from another developer.
Workers Will Need New Skills
Customizable AI could change technology careers.
Organizations may need fewer employees focused only on writing basic prompts and more people capable of preparing data, evaluating outputs, designing workflows, testing security, and connecting models with real systems.
Workers will need to understand the limits of AI rather than simply how to access it.
Important skills may include data governance, domain expertise, critical evaluation, cybersecurity, ethics, communication, and model monitoring.
Not every employee needs to become an artificial-intelligence engineer.
More workers will need enough AI literacy to recognize when a system is useful and when it is unreliable.
Education and workforce programs should avoid training people only on one company’s interface.
The deeper skill is learning how to assess and work with changing technologies.
The Release Is an Experiment as Much as a Product
Inkling enters a crowded market.
Open-weight models are already available from several organizations in the United States and abroad.
Thinking Machines must show that its system is reliable, affordable, adaptable, and worth adopting.
The company’s reputation and funding guarantee attention.
They do not guarantee long-term success.
Developers may choose competing models that are smaller, cheaper, faster, or better documented.
Businesses may remain with closed systems because they prefer managed services and established support.
Researchers may welcome Inkling’s openness but question its training data or computational requirements.
The model’s real significance will emerge through use.
The most important evidence will not be the launch announcement.
It will be what developers build, what organizations learn, and whether the system produces enough practical value to justify its cost.
Key Takeaways
Thinking Machines Lab officially released Inkling in the United States on July 15, 2026. The announcement entered the July 16 news cycle in Japan because of the time difference.
Inkling is the U.S.-based company’s first major foundation model and is available under an open-weights approach.
The model accepts text, image, and audio inputs and generates written responses.
Inkling contains approximately 975 billion total parameters, with about 41 billion activated for a particular task through its mixture-of-experts design.
Thinking Machines does not claim Inkling is the strongest overall AI model. It emphasizes customization, multimodal capability, controllable reasoning, and fine-tuning through its Tinker platform.
Open weights give developers more control but also transfer responsibility for security, safety, infrastructure, and governance.
Inkling reflects a broader shift from universal chatbots toward specialized AI systems adapted for particular organizations and industries.
The model’s long-term significance will depend on its real-world reliability, cost, accessibility, and usefulness rather than its parameter count alone.
FAQ
What is Inkling?
Inkling is a general-purpose, multimodal artificial-intelligence model developed by the U.S.-based Thinking Machines Lab.
When was Inkling released?
The official model card lists July 15, 2026, as the release date in the United States. The announcement was part of the July 16 news cycle in Japan.
Who founded Thinking Machines Lab?
The company was founded by Mira Murati and other artificial-intelligence researchers. Murati previously served as chief technology officer of OpenAI.
What does open weights mean?
It means the model’s learned numerical parameters are available for outside developers to download and modify under the applicable license. It does not necessarily mean every dataset and part of the training process is publicly disclosed.
Can Inkling understand images and audio?
Yes. It accepts text, image, and audio inputs and generates text responses.
How large is Inkling?
The model has approximately 975 billion total parameters, with about 41 billion active during a particular task.
Can Inkling run on a regular laptop?
The full model generally requires specialized computing infrastructure and is not designed to run easily on an ordinary consumer laptop.
Is Inkling better than every other AI model?
No. Thinking Machines itself says Inkling is not the strongest overall model available. Its primary advantages are intended to be flexibility and customization.
Can schools use Inkling?
Schools could potentially build tools using it, but they would need to address privacy, accuracy, student safety, academic integrity, accessibility, cost, and human oversight.
Does open AI mean risk-free AI?
No. Open-weight models can support innovation but may also be modified or misused. Developers and organizations remain responsible for how they deploy them.
Final Thoughts
Inkling is not important simply because it contains hundreds of billions of parameters.
The deeper story is control.
Most people and organizations currently use artificial intelligence through systems operated by a few large companies.
They can ask questions and receive answers, but they have limited influence over how the underlying model was designed or how it behaves.
Thinking Machines Lab is betting that the next major phase of AI will involve systems that organizations can adapt for themselves.
That could create better tools for research, medicine, education, manufacturing, finance, accessibility, and many other fields.
It could also create thousands of customized systems that are difficult to monitor and inconsistent in their safety.
Open access does not remove responsibility.
It increases the number of people who share it.
Inkling’s release gives developers another powerful foundation on which to build.
Whether that leads to useful innovation or simply another expensive model will depend on what happens after the launch.
The most successful AI technology may not be the system that performs best on a public leaderboard.
It may be the one that organizations can adapt responsibly, evaluate honestly, and use to solve problems that actually matter.
Related Articles
A New $500 Million Initiative Aims to Prepare Workers for the AI Era
https://www.newtoed.com/view-blog/a-new-500-million-initiative-aims-to-prepare-workers-for-the-ai-era-6a3e4cc4b023e
Google’s SensorFM Shows How Wearable Health Data Could Become Smarter
https://www.newtoed.com/view-blog/googles-sensorfm-shows-how-wearable-health-data-could-become-smarter-6a506795336f3
Sources
Thinking Machines Lab — Introducing Inkling
Thinking Machines Lab — Inkling Model Card
Thinking Machines Lab — Inkling
Thinking Machines Lab — Tinker Fine-Tuning Platform
WIRED — Thinking Machines Lab Releases Its First Model
Axios — Mira Murati’s Thinking Machines Debuts First AI Model
TechCrunch — Thinking Machines Releases Its First Open Model, Inkling
