Sakana AI is testing a different path for building capable AI systems: instead of training every new model from the ground up, the Tokyo-based startup is using evolution-inspired search to combine existing models into new ones.
The method, called Evolutionary Model Fusion, is designed to automatically discover useful combinations from a large pool of open-source models. The early results point to a practical question for the next phase of AI development: how far can model builders get by recombining what already exists?
A model factory inspired by natural systems
Sakana AI’s core idea is to borrow principles from evolution and collective intelligence. Its broader goal is a machine that can generate customized AI models for user-defined application domains, reducing the need to build a new model manually each time.
That ambition sits behind Evolutionary Model Fusion. The method searches through many possible ways to combine models with different capabilities, then selects promising results through an evolutionary algorithm.
The startup describes two main forms of combination. One recombines layers from different models in flow space. The other remixes model weights in parameter space. Together, these techniques give the system a large search area where unusual model combinations can emerge.
The important point is not simply that models are merged. It is that the search process is automated. Sakana AI argues that the evolutionary algorithm can find new and unintuitive solutions that conventional methods and human intuition may miss.
What Sakana AI has released
To test the approach, Sakana AI automatically developed two Japanese foundation models: a Japanese Large Language Model with mathematical capabilities and a Japanese Vision Language Model.
The company has released three Japanese foundation models:
- EvoLLM-JP, a large language model.
- EvoVLM-JP, a vision language model.
- EvoSDXL-JP, an image generation model marked as coming soon.
The releases matter because they show Evolutionary Model Fusion being applied across more than one model type. Sakana AI is not presenting the technique as only a language-model method. It is also exploring vision-language models and diffusion models for image generation.
For image generation, Sakana AI reports preliminary work on a Japanese SDXL model described as high-quality and lightning-fast. The source states that this model uses only four diffusion steps.
Early benchmark results are notable
The initial results are the main reason the work is drawing attention. Sakana AI’s automatically developed Japanese LLM and Japanese VLM achieved state-of-the-art results on several LLM and vision benchmarks, despite not being explicitly optimized for those benchmarks.
The most striking comparison involves model size. The 7-billion-parameter Japanese mathematical LLM outperformed some previous 70-billion-parameter Japanese SOTA LLMs on a number of Japanese LLM benchmarks.
Sakana AI believes the experimental Japanese mathematical LLM is good enough to be used as a general purpose Japanese LLM. That is a significant claim for a model created through automated fusion rather than a conventional development route.
The Japanese LLM is also reported to handle culture-specific content especially well. It achieved excellent results on a Japanese dataset of image-text pairs, which suggests that the fusion process may preserve or create useful Japanese-language and Japanese-context capabilities.
These results should still be read as early evidence, not a final verdict. The source describes the initial results as promising and the Japanese mathematical LLM as experimental. But the benchmark performance gives Sakana AI a concrete proof point for its nature-inspired research direction.
Why model fusion could matter
Evolutionary Model Fusion arrives at a time when open-source AI models already cover many capabilities. Sakana AI’s approach tries to treat that growing model pool as raw material. Instead of choosing one model or training an entirely new one, the system searches for useful combinations among many candidates.
For large organizations, the possible advantage is speed and cost. Sakana AI sees the combination of neuroevolution, collective intelligence and foundation models as a promising long-term research approach. According to the startup, it could help large organizations develop custom AI models faster and more cost-effectively by using open-source models before committing massive resources to fully proprietary models.
That does not mean proprietary model development disappears. The source frames the approach as a way to leverage existing open-source models before investing heavily in fully proprietary systems. In plain terms, model fusion could become a step between adopting public models as-is and building everything internally.
It also changes the role of human model designers. If the most useful combinations are difficult to predict, then the system’s search process becomes a key part of model creation. Human intuition still matters, but it may be paired with evolutionary exploration across model architectures, layers and parameters.
The company behind the work
Sakana AI is a Tokyo-based startup founded by former Google AI experts Llion Jones and David Ha. The company is focused on generative AI models inspired by nature, with systems designed to generate text, images, code and multimedia.
The founders want to create AI systems that can be sensitive and adaptable to changes in their environment, similar to natural systems with collective intelligence. That is presented as a contrast with traditional AI models, which are often designed as immutable structures.
Jones is also known as an author of the 2017 research paper Attention Is All You Need, which introduced the transformer architecture behind many AI advances. Sakana AI has raised $30 million in seed funding from investors including Lux Capital and Khosla Ventures.
The startup aims to turn Tokyo into an AI hub similar to San Francisco’s OpenAI and London’s Deepmind. Evolutionary Model Fusion is its first major signal of how it wants to compete: by making AI model development more adaptive, automated and grounded in nature-inspired search.