ai lab

Sakana AI

Sakana matters because it offers a sustained alternative to the idea that progress comes mainly from one larger monolithic model. Its work connects evolutionary model composition, autonomous experimentation, and learned orchestration. The company is most useful to understand when those ideas are evaluated together with their limitations, not when every research prototype is described as a solved scientific or product capability.^3,9,10,7

Official site ↗Models in the Atlas →

Profile

Overview

A nature-inspired lab in Tokyo

Sakana AI is a Tokyo research company founded in 2023 by David Ha, Llion Jones, and Ren Ito. Ha had studied self-organizing systems and generative models at Google Brain and Stability AI. Jones co-authored the Transformer paper. Ito brought Japanese policy and technology-company experience. They named the company after the Japanese word for fish and organized its research around collective behavior, evolution, and systems made from multiple cooperating parts.^1,2

Evolution as model development

The first public program applied evolutionary search to model composition. Evolutionary Model Merge searched both parameter space and data-flow configurations to combine existing Japanese-language, vision-language, and image models without training each result from scratch. The work was published in Nature Machine Intelligence and released with code. It gave Sakana a distinct identity: the lab would investigate how to assemble, adapt, and evolve models rather than compete only through a larger pretraining run.^3,4

Automating the research loop

The AI Scientist extended that automation thesis from model construction to research work. The system proposed ideas, wrote and ran experiments, produced manuscripts, and simulated review. A second version removed the need for a human-written code template and used agentic tree search. Independent evaluation found a striking speed and cost advantage but weak paper quality, coding failures, and citation problems. A 2026 Nature paper made the system and its limitations part of a peer-reviewed record rather than settling the question of autonomous science.^5,6,7,9

Collective intelligence becomes a product

Sakana has since turned coordination research into products. Fugu presents a learned orchestrator behind one model API, selecting and coordinating a pool of other models. The system draws on ICLR research about learned coordination and verification, while Namazu supplies Japanese-specialized models and Sakana's enterprise work targets finance, government, and industry in Japan. This connects the lab's research theme to a business, but Fugu's strongest benchmark claims still depend heavily on Sakana's own evaluations.^10,11,12

Notable contributions

01Evolutionary Model MergeSakana introduced a method that uses evolutionary search to discover parameter and data-flow combinations of existing models, then released Japanese language, vision-language, and image examples with code.^3,4
02The AI Scientist research loopThe AI Scientist integrated idea generation, coding, experiment execution, visualization, manuscript writing, and simulated review in one open research system. Later work reduced human scaffolding, while independent evaluation documented substantial quality limits.^5,6,7,9
03Learned orchestration delivered as one model APIFugu turns model selection, task delegation, verification, and synthesis into the behavior of a trained coordinator exposed through one endpoint. The product is a notable implementation of learned orchestration, not the invention of model routing or multi-agent systems.^10,11,12

Sources · 12+

Model lab #10

As of Aug 29, 2026

Sakana is productizing a coherent collective-intelligence thesis, but independent validation still trails the pace of its claims.

The lab's research identity has survived contact with the market. Fugu turns learned coordination into an API, while The AI Scientist has progressed from a provocative preprint to a peer-reviewed account with explicit limitations. Independent researchers and Nature provide unusually useful checks on the scientific-automation story. Fugu has less outside evaluation: current analysis recognizes the architectural and strategic interest while warning that its reported frontier comparisons still need reproduction. The editorial read is therefore strong on originality, research continuity, and release momentum, but it treats autonomous science and orchestration as promising systems rather than settled capabilities.

Why this editorial standing+

Sakana retains its seed score because this ranking gives unusual research influence, category originality, and continuity from papers to public systems more weight than catalog size. Evolutionary Model Merge reached a peer-reviewed venue, The AI Scientist created an influential open research line with independently documented limits, and Fugu turns Sakana's learned-coordination work into a public API. The model-evidence component reflects inspectable papers, released systems, and a coherent Japanese model program, not proof that Sakana trained a monolithic frontier model equal to the largest labs. Fugu orchestrates a pool of other models, and its strongest launch benchmarks remain Sakana-reported rather than independently reproduced. Those constraints keep Sakana below the largest frontier labs even though its research-influence score is category-leading.^3,9,10,12,7

Model evidence: 27 / 35
Research influence: 25 / 25
Release momentum: 17 / 20
Ecosystem: 8 / 10
Trajectory: 9 / 10

Notable people

David HaCo-founder and chief executivecurrent
LJ
Llion JonesCo-founder and chief technology officercurrent
RI
Ren ItoCo-founder and chairmancurrent

Revision history1 updates +

Initial researched Sakana AI dossier
Aug 29, 2026
Connected evolutionary model merging, automated research, and Fugu while preserving independent criticism of research quality and benchmark validation.