Sakana AI

Sakana AI is a Tokyo-based research lab building nature-inspired foundation models whose headline releases — AB-MCTS, EvoLLM-JP and the AI Scientist — treat model development itself as an evolutionary search problem.

Overview

Founded in 2023 by former Google researchers, Sakana AI develops techniques where multiple models combine, mutate and compete rather than being trained as one monolith. Its Japanese-English bilingual models and the open “AI Scientist” workflow, which can draft, run and review its own research experiments, are published with weights and papers so teams can adopt them directly. The lab partners with NVIDIA and Japanese enterprises on efficient training, and everything it ships is designed to run on modest hardware instead of hyperscale clusters.

Key Features

• Evolutionary model merging (AB-MCTS) that composes specialist models at inference time
• EvoLLM-JP and EvoLLM-JP-v1 bilingual Japanese/English open-weight releases
• The AI Scientist: automated literature review, experimentation and paper drafting
• TreeQuest multi-agent search library for orchestrating model populations
• Compute-efficient training recipes aimed at consumer and enterprise GPUs
• Open papers, weights and code for every major release

Pricing

Research artifacts, model weights and libraries are released openly at no cost. Enterprise collaboration and licensing around proprietary model families are handled case by case through direct contact with the lab.

Who Is It For

ML researchers, Japanese enterprise AI teams and engineers who want to compose existing models instead of paying to retrain large ones. It is especially attractive to teams with limited GPU budgets that still need strong Japanese-language performance.

Use Cases

Teams reach for Sakana AI when they need strong Japanese-English language performance without hosting a giant single model. Research groups build on its evolutionary merging techniques, while product teams combine its open weights with proprietary data.

• Composing specialist models for bilingual Japanese-English tasks
• Automating literature review and experiment drafting with the AI Scientist
• Running TreeQuest multi-agent search over model populations

Getting Started

Everything begins at the research blog and GitHub organization: papers ship alongside weights and reference code, so a working prototype usually takes an afternoon rather than a quarter. Teams with enterprise constraints can engage the lab directly for collaboration.

Results to Expect

Practitioners report quality close to far larger models at a fraction of the inference cost, plus unusually transparent documentation of how results were achieved — helpful both for engineering decisions and internal governance reviews.

FAQ

Is Sakana AI free to use?

Yes. The published models, libraries and papers are open and free to download and build on; only bespoke enterprise engagements are commercial.

Does Sakana AI offer an API?

Not as a hosted product. Teams typically self-host the released weights or integrate the TreeQuest and merging libraries into their own serving stack.

How does Sakana AI compare to alternatives?

Where labs like Mistral or AI21 sell single large models, Sakana focuses on combining many smaller ones through search, which yields strong per-dollar performance and unusually transparent research output.