Within the Tygron Platform, new AI functionality will become available to all users on October 29, 2026. This functionality makes it easier to work with the platform, perform analyses and calculations, and quickly move from a spatial question to an analysis. However, the increasing use of AI is raising more and more questions about its environmental impact. This is precisely why we at Tygron are working on AI that is not only powerful and practical, but also significantly more energy-efficient than many of the well-known large AI models.
AI
Environmental impact
AI requires a lot of computing power and therefore also a lot of energy. In addition, the data centers on which these models run use large amounts of water for cooling. Now that AI is being applied more and more widely, attention to this environmental impact is also increasing. We share this concern, and moreover, the demand for more sustainable solutions is increasingly being explicitly raised by our customers.
That is why we are continuously working on improving the efficiency of our Tygron AI software. Recently, we implemented another update that further optimizes our AI. Tygron AI has now become a much more sustainable choice than many large frontier models, such as those from OpenAI and Anthropic. For the same query, Tygron AI consumes about ten times less energy.
How is that possible?
Large frontier models, such as ChatGPT and Claude, contain around 3,000 billion or more parameters in their neural networks. As a result, they can process an enormous amount of knowledge from sources including the internet and provide answers to a wide range of questions. However, for Tygron AI, that is not necessary. Our AI must, for example, be able to explain how to build a dashboard using TQL or how a flood is simulated in Tygron. A recipe for baking cookies in Spain might be nice, but is not relevant for this application.
Therefore, we can work with a much smaller model of about 30 billion parameters. Such models possess sufficient general knowledge to reason well and, for example, write code. This smaller size does not automatically mean that the model performs less well on relevant tasks. For instance, a model like Qwen3.8 on Arena.AI is simply in the top 10 of best coding models, even though it is significantly smaller than the large frontier models.
The challenge is then to provide this smaller model with sufficient knowledge of Tygron, so that it knows our software and working methods well. We use various techniques for this. The main source is our wiki, which contains a lot of substantive knowledge. Initially, we had the AI read this information through MCP tools. Now, we also have the AI summarize the wiki around a number of key themes, so that this knowledge can be directly provided to the model.
In addition, the AI can communicate directly with Tygron via tools. For example, the AI can work with a project via the API and check if a TQL query is valid. We can also train an existing model further with additional knowledge. In doing so, the general knowledge of the model is preserved while we add specific Tygron knowledge. In AI terms, we call this fine-tuning with a LoRA adapter.
Finally, the software can now also store knowledge temporarily. By caching information from previous steps and keeping it available within the context window, the AI can carry forward part of the knowledge acquired during a conversation to subsequent questions. As a result, the same information does not have to be repeatedly searched for or supplied.
All these points result in a small, specialized model with tools that can perfectly answer Tygron questions, without the overkill of a mega model. In short, a smaller, specialized model is better for what we need at Tygron and has about 100 times fewer parameters. Depending on the quantization, each parameter represents 4, 8, or 16 bytes of data. If you have 30 billion parameters, that means about 30 GB of RAM. For a large model with 3,000 billion parameters, that is 3,000 GB of RAM.
Carbon footprint
An average data center GPU has around 100 GB of RAM. This means that a small model can run on a single GPU, whereas for a large model, you quickly need more than 10 GPUs spread across multiple machines. A data center GPU requires about 600 watts. Every extra GPU therefore also means higher power consumption.
This is of course a bit simplified. There are also ways to use large models more efficiently, for example by combining different expert models. But in broad terms, you can say that there can be a factor of 10 difference in power consumption and that much less hardware is required, which also means a smaller carbon footprint per machine. On Arena.AI, an AI comparison website from Berkeley University, similar differences can be seen when looking at the cost per token.
Transparent in power usage
Looking to the future, we will investigate how far we can push the boundaries of small, specialized models even further and how we can make these models work together. We also want to make our power consumption more transparent. That is why we recently added a log function that allows you to request exactly how many kWh a simulation or chat consumes.
See for more information: https://support.tygron.com/wiki/AI_Rules
