Datawizz, a company building specialized artificial intelligence language models as an alternative to large general-purpose models, today announced it has raised $12.5 million in a seed funding round.

Human Capital led the round, with participation from BGV, 91VC and others. The company said it will use the fresh capital to accelerate product development, expand its team and grow its operations across the United States and Europe.

Founded in 2025, Datawizz offers specialized, cost-effective models that are high-efficiency and trained on AI requests and custom enterprise data to address challenges in the AI market, which can lead to high costs and inconsistent results.

“Companies spent $8.4 billion on API calls to LLMs in just the first half of 2025 — more than double the figure for all of 2024,” said Iddo Gino (pictured), founder and chief executive of Datawizz.

Every time a large language model is used, it processes tokens — the basic units of text such as words, word fragments, or characters — to generate input and output. Tokens are essentially the internal “currency” for LLMs. Most AI services that provide access to models via an application programming interface charge based on the number of input and output tokens consumed.

Models such as OpenAI’s GPT-5 break text into tokens to provide answers. Reasoning models, which “think” through problems with an internal monologue before responding, consume even more tokens. According to Artificial Analysis, reasoning models can use up to 20 times more tokens than standard AI models.

The same analysis found that large, generic nonreasoning models are also consuming more tokens over time. In benchmark tests, GPT-4o used 5.5 million tokens, while GPT-4.1 required 7 million, a 27% increase.

“The pace of adoption is incredible,” Gino said. “The challenge is that many applications have upside-down unit economics and deliver inconsistent results.”

Although the industry continues to rely heavily on large, general-purpose models, Gino said Datawizz aims to disrupt this trend with specialized expert models that require fewer resources while still delivering accurate answers. The company offers fully managed services that profile an organization’s AI requests, train expert models and automatically route queries to them.

Using this solution, a company could have dozens of specialized models in its stable, each about 1,000 times smaller than a regular LLM and trained to answer specific requests. Questions that don’t fit the current specialized models can still be routed to generalized LLM providers or companies can constrain all of their AI use to fine-tuned in-house models.

Datawizz reports transformative results, achieving over an 85% reduction in AI costs, a speed improvement of five to to 15 times, and more precise, predictable outcomes.

According to the company, enterprise users gain access to models fully owned by the customer, running on-premises, on-device or in the cloud. This ensures no external data sharing and eliminates the need for expensive API calls.

“Token costs for monolithic models break unit economics for many workloads,” said Brian Blond, investment partner at Human Capital, who will join Datawizz’s board. “Datawizz provides an elegant and essential solution that will become a core component of the AI stack, and we are thrilled to partner with them to build the next generation of AI.”

Datawizz predicts that by 2030, around 70% of AI traffic will be handled by small, specialized models with fewer than 5 billion parameters. For comparison, while OpenAI has never disclosed the size of GPT-4, estimates place it at around 1.8 trillion parameters. GPT-5’s scale remains unknown, though analysts estimate it could range between 2 trillion and 5 trillion.

Image: Datawizz