On July 15, 2026, Thinking Machines Lab released Inkling, its first general-purpose foundation model and one of the largest open-weights artificial-intelligence models yet made publicly available. Inkling is designed not merely as another chatbot, but as a broad, multimodal reasoning engine that companies, universities, research laboratories and software developers can download, operate, fine-tune and adapt to their own work.

Thinking Machines Lab is a San Francisco artificial-intelligence research and product company founded by former OpenAI Chief Technology Officer Mira Murati, who serves as its cofounder and chief executive officer. The founding group has included prominent former OpenAI researchers and engineers such as John Schulman, Barret Zoph, Lilian Weng, Andrew Tulloch and Luke Metz. The company’s larger mission is to develop AI systems that extend human judgment while allowing individuals and organizations to shape models around their own knowledge, objectives and working methods. (Reuters⁠)

The official Inkling announcement and technical information are available at:

https://thinkingmachines.ai/news/introducing-inkling/

The model card is available at:

https://thinkingmachines.ai/model-card/inkling/

Inkling’s importance does not come from being declared the most capable model in every benchmark. Thinking Machines explicitly acknowledges that it is not presently the strongest overall model, whether open or proprietary. Its significance lies in the combination of very large scale, native multimodal processing, controllable reasoning, a one-million-token context window, an open-weights license and an accompanying cloud platform that makes customization considerably more accessible.

What Inkling Is

Inkling is a decoder-only, multimodal Mixture-of-Experts transformer containing approximately 975 billion total parameters. However, only about 41 billion parameters are activated while processing any particular token. This distinction is essential.

A conventional dense model may use nearly all of its parameters for every token it processes. Inkling instead contains 256 specialized computational “experts.” A routing mechanism evaluates each incoming token and sends it through six selected experts, together with two shared experts that remain active. Different experts can develop greater competence in different kinds of patterns, subjects or operations. The model can therefore draw upon a very large total knowledge and capability base without performing calculations through all 975 billion parameters for every word, image element or audio segment. (Hugging Face⁠)

This sparse architecture does not make Inkling a small model. All 975 billion parameter weights must normally be stored somewhere, even though only a fraction participate in each computational step. Mixture-of-Experts design primarily reduces the arithmetic performed during inference; it does not eliminate the considerable memory, storage, networking and model-loading requirements associated with a nearly one-trillion-parameter system.

Inkling was pretrained on approximately 45 trillion tokens drawn from text, images, audio and video. It supports a context window of as many as one million tokens, although the managed Tinker service initially provides context configurations of 64,000 or 256,000 tokens. A one-million-token context could theoretically encompass several long books, a large collection of technical documentation, an extensive software repository or a substantial body of corporate records in a single working session. Actual usable context depends on the serving system, memory capacity, application design and the amount of space required for the model’s output. (Thinking Machines Lab⁠)

The model was pretrained and subsequently improved through supervised fine-tuning and large-scale reinforcement learning. Thinking Machines reports that it performed more than 30 million reinforcement-learning rollouts, exposing the model to reasoning, mathematics, coding, tool use, image interpretation, audio analysis, conversation and safety tasks. The production model was trained on NVIDIA GB300 NVL72 systems. (Thinking Machines Lab⁠)

How Inkling Processes Information

Inkling generates its output autoregressively. It examines the information already present in the context and repeatedly predicts the next appropriate token. Those tokens may represent words, programming syntax, mathematical notation, internal reasoning elements or portions of a structured response.

Its architecture includes several efficiency mechanisms. Inkling alternates sliding-window attention layers with global-attention layers at an approximate ratio of five to one. Sliding-window attention concentrates on nearby information and reduces the computational burden of repeatedly comparing every token with every other token. Periodic global-attention layers allow information to be integrated across the larger context.

Instead of relying exclusively on the rotary positional encoding used by many contemporary models, Inkling uses learned relative positional information. This helps the system identify how tokens relate to one another based on their distance and position within a sequence and is intended to support extrapolation to very long contexts.

Inkling also incorporates short one-dimensional convolutional operations that help capture local patterns. These operations can recognize relationships among nearby tokens while reserving the attention and expert layers for more complex or longer-range associations. (Hugging Face⁠)

The model includes multi-token-prediction layers that can propose several future tokens simultaneously. During deployment, those proposals can be checked by the principal model through a technique called speculative decoding. When the proposed tokens are accepted, the system can generate an answer more quickly without changing the resulting content. (Hugging Face⁠)

Native Text, Image and Audio Reasoning

Inkling is natively multimodal rather than being a text model with completely separate vision and audio systems attached afterward. Text, image and audio representations are converted into embeddings and processed jointly through the transformer.

Images are divided into patches measuring approximately 40 by 40 pixels. A lightweight hierarchical multilayer-perceptron system transforms those patches into representations the model can analyze alongside text. This enables Inkling to describe images, interpret charts and diagrams, answer questions about visual material, analyze mathematical figures and combine visual observations with code-based tools such as Python.

Audio is transformed using discretized Mel spectrogram representations. Audio segments of approximately 100 milliseconds are converted into frequency-based information and embedded for processing by the model. Inkling can transcribe speech, follow spoken instructions, answer questions about recordings and reason over longer audio passages. (Thinking Machines Lab⁠)

The training corpus also included video. The architecture can represent a temporal sequence of image information, but Thinking Machines and Hugging Face caution that Inkling’s out-of-the-box video performance has not yet been comprehensively evaluated. It should therefore be viewed as having an architectural foundation for video applications rather than as a fully validated general-purpose video-understanding system. (Hugging Face⁠)

Controllable Reasoning Effort

One of Inkling’s most useful characteristics is its controllable thinking effort. Developers can select settings ranging from no explicit reasoning through minimal, low, medium, high, extra-high and maximum effort.

A low-effort response may be appropriate for classifying an email, extracting a customer number or answering a routine factual question. A high-effort response may be preferable for debugging software, analyzing an engineering problem, evaluating competing business strategies or coordinating a complicated sequence of tools.

Reasoning consumes output tokens, processing time and energy. Allowing an application to vary the amount of reasoning therefore creates an economic control. A company can reserve extensive reasoning for decisions that justify the added expense while processing routine work with lower latency and lower token usage.

Thinking Machines reports that Inkling can achieve a given level of performance on some agentic coding tests while using substantially fewer reasoning tokens than certain competing open models. Such comparisons are benchmark-dependent and should be independently tested for each real application, but the design objective is important: the model is being optimized not simply to reason, but to reason efficiently. (Thinking Machines Lab⁠)

What Inkling Can Accomplish

Inkling is intended to be a general foundation model rather than a narrowly specialized product. Its demonstrated capabilities include text generation, document analysis, question answering, mathematical reasoning, software development, instruction following, tool use, agentic task execution, image interpretation, chart analysis, speech transcription, audio question answering and probabilistic forecasting.

A customized deployment could serve as a software-engineering agent that reviews repositories, writes tests, investigates failures and proposes repairs. It could become a research assistant capable of processing scientific articles, laboratory notes, diagrams and recorded meetings. It could analyze contracts, manuals, regulatory records and internal policies. It could operate as a technical-support system that listens to a customer, views an equipment photograph and searches relevant service documentation.

The model’s long context could also make it useful for work requiring integration across large collections of related material. An engineering organization might supply requirements, schematics, test results, source code and operating logs. A manufacturer might combine maintenance manuals, sensor records, quality reports and audio descriptions from technicians. A financial organization might provide policies, research, transaction documentation and market records, subject to appropriate security and regulatory controls.

Inkling’s agentic capabilities allow it to do more than produce text. When connected to authorized tools, an Inkling-based system could search databases, execute code, call business applications, manipulate files, perform calculations or coordinate a sequence of actions. The model itself does not automatically receive unrestricted access to these systems; the organization must construct a controlled tool layer with identity management, permissions, logging, validation and human approval for consequential actions.

Open Weights Versus Open Source
 Inkling is described as an open-weights model and is released under the Apache 2.0 license. Open weights means that the learned numerical parameters are available for downloading, inspection, operation and modification. This permits an organization to host the model independently, retain greater control over its data and create customized derivatives.

The term does not necessarily mean that every element of the training process is completely reproducible. Rebuilding Inkling from the beginning would require the original training corpus, filtering methods, data mixtures, training software, reinforcement-learning environments, evaluation systems and enormous quantities of computing power. The available weights nevertheless provide substantially greater operational freedom than a closed model accessible only through its developer’s application-programming interface. (Thinking Machines Lab⁠)

How Inkling Can Be Customized

Inkling can be customized at several levels.

The simplest customization uses system instructions, prompt templates, tool definitions, retrieval systems and structured workflows without changing the model weights. This approach is fast and appropriate when the organization primarily needs the model to follow its terminology, retrieve approved information and interact with existing applications.

Retrieval-augmented generation can connect Inkling to a private knowledge repository. Documents are indexed, relevant passages are retrieved for each request and the passages are inserted into the model’s context. This permits information to be updated without repeatedly retraining the model.

Fine-tuning changes the model’s behavior more deeply. Thinking Machines makes Inkling available through Tinker, its managed training application-programming interface. Tinker allows developers to supply supervised examples or reinforcement-learning environments while Thinking Machines manages the distributed GPU infrastructure, scheduling, checkpointing and resource allocation. (Thinking Machines Lab⁠)

Tinker uses Low-Rank Adaptation, commonly called LoRA. Rather than modifying every one of Inkling’s hundreds of billions of weights, LoRA trains comparatively small adapter matrices that influence the model’s operation. The original model remains intact, while the adapter teaches it a particular style, workflow, domain behavior or decision pattern. Thinking Machines says properly configured LoRA training can approach the learning performance of full fine-tuning while requiring less computation and giving users greater flexibility. (Thinking Machines Lab⁠)

Reinforcement learning provides another level of adaptation. Instead of supplying only examples of correct answers, an organization can construct an environment in which the model performs a task and receives a reward based on the outcome. A coding model might receive credit when software passes tests. A logistics agent might be rewarded for meeting delivery objectives while controlling cost. A scientific agent might be evaluated according to the quality of a proposed experimental sequence.

Customizations That Would Make Practical Sense

A healthcare organization could customize Inkling around medical terminology, clinical documentation formats, scheduling procedures and internal policies. It would still require professional oversight, validation and strong privacy controls and should not be presumed to replace licensed clinical judgment.

A manufacturer could fine-tune the model on equipment documentation, quality procedures, maintenance records and diagnostic workflows. A multimodal version could allow a technician to photograph a machine component, describe an abnormal sound and obtain a troubleshooting procedure grounded in the correct manual.

An insurer could customize it around underwriting guidelines, claims documentation and regulatory requirements. A bank could train it to interpret internal policies, identify incomplete documentation and assist analysts with investigations while keeping final authority with accountable employees.

A software company could create a model that understands its architecture, coding standards, development tools and testing practices. A law firm could customize document analysis and drafting behavior for its specialties, provided it maintains confidentiality controls and attorney review.

A university or research laboratory could adapt Inkling to scientific literature, simulation tools, experimental records and specialized mathematical notation. Government agencies might customize it around statutes, regulations, case-management systems and public-service workflows while operating it within a controlled environment.

Customization is most valuable when an organization possesses distinctive knowledge, repeated work processes and clear methods for evaluating success. A model should not be fine-tuned merely because customization is available. The organization needs high-quality training examples, measurable evaluation criteria, permission to use the underlying data and a maintenance process that identifies when policies or facts have changed.

Advantages of Customization

A properly customized model may become more accurate within the organization’s domain, follow internal procedures more reliably, use approved terminology and generate outputs in required formats. It may require shorter prompts because recurring expectations have been incorporated into the adapter or workflow.

Self-hosting can also improve data control. Sensitive information can remain within the organization’s selected cloud environment, private data center or sovereign-computing region. Open weights reduce dependency on a single application-programming-interface provider and allow the organization to preserve a validated version rather than having its underlying model changed without notice.

Customization can improve productivity by reducing the amount of supervision required for repetitive work. However, the benefit depends on disciplined evaluation. Fine-tuning on poor or biased examples can make a model consistently reproduce the same errors. Organizations need benchmark suites based on their own work, adversarial testing, security reviews, human escalation policies and ongoing monitoring.

Where Inkling Runs

Inkling can operate in three broad ways.

First, organizations can use managed cloud inference. Hugging Face makes Inkling accessible through participating inference providers, allowing developers to submit requests without buying or operating the underlying hardware. Tinker similarly provides managed infrastructure for fine-tuning. For most small and medium-sized organizations, this is likely to be the fastest and economically most practical starting point. (Hugging Face⁠)

Second, enterprises can deploy the weights in their own cloud accounts or data centers using inference frameworks such as Hugging Face Transformers, SGLang or vLLM. These frameworks can divide the model across multiple GPUs and expose an application-programming interface for internal products.

Third, heavily quantized versions can be operated through llama.cpp and related local-inference tools. Quantization reduces the number of bits used to represent each parameter, dramatically lowering memory requirements. The trade-off can include lower accuracy, slower processing when weights are moved from system memory, reduced multimodal functionality or other limitations depending on the quantization method.

Hardware Requirements for the Full Model

Inkling is far too large to run in full quality on an ordinary desktop computer or a single conventional server.

According to Hugging Face, the full BF16 checkpoint requires approximately two terabytes of GPU memory. It is intended for NVIDIA Hopper-generation or later accelerators. Two terabytes of aggregate GPU memory would require, for example, a large multi-GPU, multi-node system; even 16 GPUs with 141 gigabytes each would provide only modest memory headroom after accounting for runtime overhead, communication buffers and the key-value cache used for context. (Hugging Face⁠)

The NVFP4 version requires approximately 600 gigabytes of GPU memory and is intended for NVIDIA Blackwell hardware. Eight GPUs with roughly 80 gigabytes of usable memory each might hold the static weights in principle, but production operation would require additional capacity for the context cache, batching, runtime overhead and multimodal processing. A practical high-throughput or long-context deployment may therefore need eight high-memory Blackwell GPUs or multiple interconnected servers, depending on the particular accelerator configuration. (Hugging Face⁠)

The one-million-token context is especially demanding. Although the model weights remain fixed, the key-value cache grows as the context becomes longer and as more users are served simultaneously. An organization that only needs 32,000 or 64,000 tokens can cap the context and substantially reduce memory requirements. Hugging Face specifically advises limiting the maximum model length when key-value-cache memory becomes a constraint. (Hugging Face⁠)

Production performance also depends on high-bandwidth GPU interconnection. A model divided across many accelerators must repeatedly exchange intermediate results. NVIDIA NVLink, NVSwitch and high-speed InfiniBand or Ethernet networking can therefore be nearly as important as total GPU memory. Deploying the full model is a data-center engineering project, not simply a matter of installing software on a powerful workstation.

Inkling-Small: The Lighter Version

The term used by Thinking Machines is Inkling-Small. It is not merely a compressed copy of the main model. It is a separately trained Mixture-of-Experts model containing approximately 276 billion total parameters and 12 billion active parameters, compared with Inkling’s 975 billion total and 41 billion active parameters.

Inkling-Small was trained with a similar architecture and post-training system but benefited from improvements in the pretraining data and training recipe. As a result, Thinking Machines reports that the smaller model matches or exceeds the full model on a number of benchmarks. That does not mean it is universally equal. The larger model is expected to retain advantages on some difficult, broad or uncommon tasks, but the smaller model may deliver a better combination of capability, latency and operating cost for many applications. (Thinking Machines Lab⁠)

As of July 16, 2026, Inkling-Small remains a preview. Thinking Machines says it is completing testing and will release the full weights after that work is finished. Tinker lists support for Inkling-Small as “coming soon.” Consequently, exact production hardware requirements and pricing should not yet be treated as finalized. (Thinking Machines Lab⁠)

Estimated Hardware Requirements for Inkling-Small

Because official deployment-memory figures have not yet been published, its hardware requirements must be estimated from the announced parameter count.

At 16-bit precision, 276 billion parameters require approximately 552 gigabytes merely to store the raw weights. Allowing for runtime overhead and context memory, a practical BF16 installation might require eight or more GPUs with approximately 80 gigabytes each, or a smaller number of newer accelerators with substantially greater memory.

At eight-bit precision, raw weight storage would be approximately 276 gigabytes. At four-bit precision, it would be approximately 138 gigabytes before overhead. A four-bit Inkling-Small deployment could therefore potentially fit across two 80-gigabyte GPUs for restricted contexts and light workloads, although three or four accelerators would provide more practical space for runtime overhead, multimodal inputs, longer contexts and concurrent users.

Aggressive two-bit or one-bit quantization might reduce the model enough to operate from a high-memory workstation or CPU server, but this would not necessarily provide the speed or quality expected from a fully functional production deployment. The active parameter count affects computation, but the entire set of experts must remain accessible in GPU memory or be transferred from system memory or storage. Offloading makes local operation possible on less sophisticated hardware but can introduce severe latency.

For an enterprise seeking reliable, responsive operation, Inkling-Small will most likely be attractive on a modest multi-GPU server or through a cloud inference service rather than on an ordinary personal computer.

How the Smaller Model Could Reduce Cost

Inkling-Small activates 12 billion parameters per token, about 29 percent of the 41 billion active parameters in the full model. Its total stored parameter count is also approximately 28 percent as large. These proportions do not translate directly into an exact 72 percent cost reduction because inference cost also includes attention, communications, memory access, context processing, batching and system overhead. Nevertheless, the architecture indicates that Inkling-Small should require substantially less memory and computation.

A smaller deployment may use fewer accelerators, consume less electricity, occupy less data-center space and produce answers with lower latency. It may also serve more simultaneous users on a given amount of hardware. Cloud providers could therefore offer it at a considerably lower price per token than the main model.

For routine enterprise work—document routing, information extraction, customer assistance, internal search, standard software tasks or structured report generation—the small model may be economically superior even when the larger model produces somewhat stronger answers on the most difficult questions.

A practical system could also route work between the two models. Inkling-Small could handle routine requests, while complex, uncertain or consequential problems would be escalated to full Inkling. This model-routing architecture could preserve high capability while controlling overall operating cost.

Why Inkling Could Become Widely Adopted

The most accurate term is probably “viable,” although Inkling could also become “viral” among developers if its open release produces a rapid ecosystem of tools, fine-tunes and demonstrations.

Several factors could make it viable. It combines open weights, multimodal reasoning, long context, agentic capability and controllable reasoning in one model. The Apache 2.0 license permits broad commercial use. Day-one support from Hugging Face, Transformers, SGLang, vLLM and llama.cpp lowers the technical barrier to experimentation. Tinker gives organizations a path to fine-tune the model without first constructing a distributed training cluster. (Hugging Face⁠)

The full model is too large for casual local deployment, which may limit its immediate use. However, cloud inference, quantized releases and Inkling-Small could expose the architecture to a much wider group of developers. Open-weight models often acquire value through an ecosystem of adapters, quantizations, inference optimizations, evaluation results and specialized versions contributed by organizations other than the original developer.

Inkling also arrives at a time when many companies want more control over proprietary data, model behavior and vendor dependence. An organization can validate a particular Inkling checkpoint, host it within its own security boundary and create adapters that it owns. That is fundamentally different from building an important business process around a remotely controlled model that may change in ways the customer cannot inspect.

Jobs an Inkling-Based System Could Perform

An Inkling deployment could function as a coding assistant, research analyst, document-processing agent, multimodal technical-support specialist, forecasting assistant, quality-control analyst, compliance aide, engineering knowledge system or internal enterprise copilot.

It could inspect text and photographs in an insurance claim, listen to a recorded customer explanation, compare the information with policy documentation and prepare a structured file for a human adjuster. It could review a factory maintenance report, interpret an image of the equipment, consult the correct service manual and generate a proposed diagnostic sequence. It could analyze a software issue, inspect the repository, execute tests and prepare a patch for developer approval.

These systems should generally be viewed as tools for extending accountable employees rather than as unsupervised replacements for professional judgment. The greatest productivity gains are likely to arise when AI completes information-intensive portions of a job while employees define objectives, validate evidence, resolve ambiguity and remain responsible for final decisions.

A Broader Expansion of AI Choice

Inkling is part of a larger transition from a market dominated by a small number of general-purpose proprietary models toward a more diverse ecosystem of closed commercial models, open-weight foundation models, compact local models, reasoning models, multimodal models and highly specialized enterprise systems.

This abundance creates a genuine strategic choice for companies and public organizations. They can select a premium cloud model for maximum general capability, an open-weight model for control and customization, a compact model for low-cost local processing or a coordinated combination of models that routes each assignment to the least expensive system capable of performing it reliably.

Competition among model families should improve capability, reduce inference costs, accelerate hardware optimization and weaken dependence on any single vendor. It will also allow organizations to match artificial intelligence more precisely to the work being performed rather than forcing every application through one general-purpose service.

The ultimate economic value will not come simply from having more chatbots. It will come from incorporating appropriate models into engineering, manufacturing, healthcare administration, logistics, software development, scientific research, finance, education and government operations. When models can be customized around an organization’s knowledge and evaluated against its real objectives, they can reduce the time employees spend locating information, preparing routine material, checking work and coordinating complex processes.

Inkling illustrates this emerging direction. It is an enormous model, but its most important idea is decentralization: giving organizations access to a capable foundation they can adapt, control and make their own. Inkling-Small extends that proposition toward lower-cost hardware and broader deployment. Together with the growing number of competing open and proprietary models, these systems are creating a rapidly expanding menu of AI capabilities that can raise productivity throughout the economy—provided organizations combine the technology with sound data, rigorous evaluation, security controls and accountable human judgment.