Sakana AI optimizes model performance through evolutionary merging

Sakana AI uses evolutionary techniques to discover optimal ways to combine different open-source models. This method, called Evolutionary Model Merge, works with the collective intelligence of existing models on platforms like Hugging Face, which hosts over 500,000 models across various modalities. The company’s approach targets the development of small, specialized AI systems that occupy specific ecological niches rather than relying on a single, massive foundational model.

The company released several Japanese foundation models using this evolutionary process. Their 7B parameter Japanese Math LLM, EvoLLM-JP-v1-7B, achieved top performance on various Japanese LLM benchmarks and exceeded the performance of some previous 70B parameter Japanese LLMs. The EvoVLM-JP-v1-7B model handles culturally-specific content well and reaches top results on Japan-sourced datasets of Japanese image-description pairs. Sakana AI also developed EvoSDXL-JP, a high quality, lightning fast 4-diffusion-step Japanese-capable SDXL model.

The evolutionary approach requires relatively little compute resources because it does not rely on gradient-based training. The method discovers novel ways to merge models from different domains, such as non-English language and Math, which human experts might struggle to identify. While the method works without backpropagation, the company notes that they could employ gradient-based backpropagation to improve performance further.

CycleQD automates agentic research workflows

Sakana AI is developing technology to automate the creation of foundation models for user-specified application domains through the CycleQD framework. This method evolves a swarm of LLM agents, each with its own niche, to handle difficult agentic workflow tasks. CycleQD uses model merging as an evolutionary crossover operation and Singular Value Decomposition (SVD) as a mutation operation. The framework uses a Quality Diversity-based selection operation to ensure the population of models remains diverse and high performing.

In experiments training Llama3-8B-Instruct on computer science tasks, CycleQD outperformed traditional fine-tuning and model merging approaches. The model achieved high pass@1 scores on the MBPP coding benchmark and high success rates on Database (DB) and Operating System (OS) tasks. This approach uses specialized expert LLMs for each skill to add high-quality models to the collection. The method focuses on improving one skill while using other skills as behavior characteristics to maintain balance.

This specialized agentic approach competes with existing research agents like OpenAI’s Deep Research. Microsoft also provides a Researcher agent that uses OpenAI’s Deep Research model and Microsoft 365 Copilot’s orchestration to pull context from emails, meetings, and files. While these tools create reports, Sakana AI aims to automate the very machinery that generates the models themselves. Will these specialized swarms eventually replace the need for generalist frontier models?

Strategic funding and the intelligence race

Sakana AI secured a $135M Series B round in November 2025, which set a post-money valuation of $2.65B. Investors in this round included MUFG, Khosla Ventures, Macquarie, New Enterprise Associates, and In-Q-Tel. Earlier funding rounds included a $114M Series A in June 2024 and a $30M Seed round in January 2024. The company works closely with Nvidia on joint research and the development of AI infrastructure in Japan.

Japan seeks to build its own AI ecosystem to reduce dependence on foreign technologies. Sakana AI’s focus on energy-efficient, collaborative models suits Japan’s need for sustainable AI solutions. The startup’s leadership includes David Ha, who helped establish the Google Brain research team in Tokyo, and Llion Jones, a co-author of a seminal paper in the field.

The rise of independent AI development coincides with increased competition between US and Chinese technology. While Microsoft develops its own foundation models like MAI-Base-2 to gain independence from OpenAI, Sakana AI focuses on the Japanese market. The company’s strategy relies on many smaller models working together instead of one massive system. This distinction separates Sakana AI from the scaling trends seen in models like the 2.8 trillion parameter MAI-Base-2.