OrganizationCompany
Sakana AI
Sakana AI is a Tokyo research and development company founded in July 2023 by David Ha, Ren Ito and Llion Jones, known for evolutionary model merging, projects such as The AI Scientist, and Japanese-language models named Namazu.
Founded 2023Website sakana.ai Era Generative AI for everyone
A Tokyo research company
Sakana AI is a research and development company based in Tokyo, founded in July 2023.[1] Its registered name is Sakana AI K.K. and its headquarters is in the Azabudai Hills Mori JP Tower in Minato-ku, Tokyo.[1] The company's own overview splits its work into research output, a product line and a family of Japanese-language models.[1]
Founders and leadership
The three founders hold the top roles: David Ha is chief executive, Ren Ito is chairman and Llion Jones is chief technology officer.[1] Jürgen Schmidhuber serves as chief scientific advisor and works with the company's RSI Lab on recursive self-improvement and physical AI.[1] Recursive self-improvement here means AI systems taking part in building the next generation of AI systems; physical AI means models that act through robots and other hardware rather than only in software. Ai Fujiki is chief financial officer; she joined in July 2025 as head of corporate planning after about ten years at Goldman Sachs.[1]
Research: merging models instead of training them
The company lists The AI Scientist, the Darwin Gödel Machine, ShinkaEvolve and AB-MCTS among its research output.[1] The line of work its public materials describe in most detail is model merging: rather than training one large model from scratch, take several open models that already exist and combine them into a single model.
On 19 March 2024 Sakana posted the paper behind EvoLLM-JP, which describes an evolutionary search over combinations of open models and presents the method as working without extra training data or compute.[2] The released 7B model merges Shisa Gamma 7B v1, WizardMath 7B v1.1 and Abel 7B 002, and its weights went up on 6 March 2024.[3]
Merging is interesting because open models tend to be good at different things: they were trained on different data with different objectives, so their strengths are complementary. The basic operation is to average or otherwise mix their weights in the hope that the result inherits strengths from several of them. Evolutionary search turns this into a search problem: instead of knowing in advance which models to combine and in what proportion, generate many combinations, score them, keep the better ones and recombine those into the next generation.
Products
The product line includes Sakana Chat, Sakana Marlin and Sakana Fugu, while the Japanese models are named Namazu.[1] The overview lists the products and the Japanese models separately, so they read as two distinct lines rather than one.[1]
Investors
The company's latest disclosed round is a Series B. Japanese investors listed include ITOCHU, ANA, KDDI, NEC, Fujitsu, MUFG and Mizuho, and overseas investors listed include Google, NVIDIA, Khosla Ventures, NEA, Lux Capital, Salesforce Ventures and In-Q-Tel.[1] The list spans trading houses, an airline, telecom and electronics manufacturers and banks alongside corporate and venture investors from abroad — broad coverage for a young company.
Where it sits
Sakana AI is a young company that both publishes research and ships products under its own name, with Japanese-language models on a separate line.[1] Judging by the projects it lists, its research leans toward getting more out of models that already exist and toward automating parts of research itself, and its RSI Lab works on recursive self-improvement.[1] The Series B shows interest from both Japanese industry and overseas technology investors.[1]
CapabilitiesLarge language models
TechniquesFoundation models
Ecosystem & regionEurope
HistoryResearch history
20241
-
ReleaseSakana AI merges open models by evolutionary search to build a Japanese LLM
On 19 March 2024, Sakana AI posted a paper on using evolutionary search to merge open models without extra training data or compute, and released a 7B Japanese model that also handles mathematics as an experiment.
Models and products over time
1-
2024
-
Mar 19, 2024
EvoLLM-JP
Model
EvoLLM-JP is a Japanese model from Sakana AI built by evolutionarily searching merges of three open-source models instead of training from scratch; the 7B result has mathematics ability none of its sources had alone.
-
Mar 19, 2024
EvoLLM-JP
Model
People1
Founders1
Notes and sources
- About | Sakana AIhttps://sakana.ai/company-info/
- Evolutionary Optimization of Model Merging Recipeshttps://arxiv.org/abs/2403.13187
- SakanaAI/EvoLLM-JP-v1-7Bhttps://huggingface.co/SakanaAI/EvoLLM-JP-v1-7B
Cite this page: Sakana AI. TellWorld, https://www.tellworld.net/companies/sakana-ai. Updated 2026-11-08