The global race for Artificial Intelligence has entered a new phase. If until a few months ago the attention was focused on models that achieved the best performance on text, now the focus has shifted towards systems that can understand and reason simultaneously on text, images, audio and video. In this context, the company Thinking Machines Lab, founded by Mira Murati, has introduced Inkling, a giant model with about 1 trillion parameters, which is published as an “open” model and promises to be one of the most advanced multimodal systems available to researchers and developers. Although the technical figures are impressive, the importance of Inkling does not lie only in its size. It is related to the philosophy of construction and the direction in which the Artificial Intelligence industry is moving. But what makes this system really special and how can it change the race in AI?
FROM CHATBOTS TO SYSTEMS THAT UNDERSTAND THE WORLD
The first generation models were built primarily for text. Later, the capabilities to analyze images and, more recently, audio were added. Inkling is designed so that these modalities are not treated as separate add-ons, but as part of the same reasoning process. In practice, this means that the model can analyze a photo, listen to an audio recording, and combine that information with text to produce a single response. This is an important step towards systems that don’t just “read” but interpret complex situations in a similar way to humans.
1 TRILLION PARAMETERS DOESN’T MEAN ALL ARE USED
One element that can be confusing is the number of parameters. Inkling has about 975 billion parameters, but only about 41 billion are activated during each response. This is achieved through the Mixture of Experts (MoE) architecture. Instead of the entire model working on each question, the system selects only the “experts” it needs for that task. It is similar to the logic of a hospital: not every doctor deals with every patient, but the appropriate specialist is called. The result is a very large model, but more efficient in the use of resources.
A CONTEXT WINDOW OF 1 MILLION SYMBOLS
One of the most important features is support for a context of one million symbols. In practical terms, this means that the model can remember and analyze an extremely large amount of information within a conversation or document. This opens up new possibilities for the analysis of very long contracts, legal documentation, scientific research, financial analysis, historical archives, or the development of autonomous agents working on complex projects. For many companies, this may be more important than increasing performance in classic AI tests.
A DIFFERENT ARCHITECTURE FROM MOST Latest MODELS
Inkling doesn’t quite follow the technical path of other popular models. Instead of the usual RoPE mechanism for positioning information, it uses a Relative Attention system, which aims to better preserve the relationships between elements in very long documents. It also combines two forms of attention:
– Global Attention, where the model sees the entire context;
– Sliding Window Attention, where it focuses only on the closest parts.
This hybrid approach significantly reduces computational costs without losing the quality of reasoning. Another interesting element is the use of a Short Convolution (SConv), a layer that helps the model understand local relationships between words or elements without burdening the main attention mechanism.
TRAINING OVER 45 TRILLION SYMBOLS
Thinking Machines says Inkling has been trained on around 45 trillion symbols, including text, images, audio, and video. This is an incredibly large amount of data, which is intended to give the model better generalization abilities across a variety of tasks. However, as with any generative model, real-world performance will depend on independent testing by the scientific community.
AN “OPEN” MODEL, BUT NOT FOR EVERY COMPUTER
Although Inkling is released as an open source model, this does not mean that it can be easily installed by any user. The standard version requires about 2 terabytes of graphics memory (VRAM). Even the most compressed version requires about 600 GB of VRAM, a capacity that is only found in large data centers or cloud infrastructure. In practice, universities, research laboratories and large companies are the main users of this model. For individual developers, the alternative remains use through cloud platforms or optimized versions.
WHY IS IT IMPORTANT FOR THE INDUSTRY?
Inkling’s release marks a significant shift in the AI race. Instead of the competition being focused solely on chatbots, companies are building models that can understand the world through multiple types of information at once. This is particularly important for robotics, autonomous vehicles, medical diagnostics, video analytics, cybersecurity, and industrial automation. At the same time, releasing as an open model could tip the balance between companies that keep their technology private and the growing open-source community.
WHAT IS LEFT TO BE SEEN?
Inkling presents an ambitious architecture and a novel combination of techniques that promise efficiency and advanced multimodal capabilities. However, its real impact will not be measured by the number of parameters or architectural innovations, but by how it performs in practical use, how easily it can be adapted for specific applications, and whether it is widely adopted by the developer community.
If these expectations are confirmed, Inkling could mark one of the most important steps towards the next generation of Artificial Intelligence models, where text, image, and voice are no longer treated as separate systems, but as part of the same intelligent process.