Sakana AI
Sakana AI is a Tokyo research company founded by David Ha, Ren Ito and Llion Jones. Its work includes model merging, distillation and new neural architectures. Namazu is its Japanese business model; many earlier experiments have downloadable research weights.
Company & leadership
- CEO
- David Ha
- Chairman
- Ren Ito
- CTO
- Llion Jones
- Legal entity
- Sakana AI K.K.
- Ownership and funding
- Venture-backed company; the April 2026 Series B update reports ¥32 billion raised in that round.
Founders
- David Ha, Ren Ito and Llion Jones
The founders brought research and business experience
David Ha led Google Brain's Tokyo research team. Llion Jones co-authored the Transformer paper. Ren Ito worked in Japan's foreign ministry and held executive roles at Mercari and Stability AI before helping establish Sakana. The current leadership page also names Jürgen Schmidhuber as Chief Scientific Advisor.
Its research develops models through several methods
Sakana's early work used evolutionary methods to combine existing models. Later projects studied distillation, model self-improvement and different neural architectures. Its Series B statement connects this work with enterprise deployment in Japan, including partnerships with financial institutions.
The Series B grew after its first announcement
The company announced the round in November 2025 and updated it in April 2026. The revised total is ¥32 billion, reported as US$200 million. Named investors include MUFG, Khosla Ventures, Google, Salesforce Ventures and Citi. This is a funding-round total, not annual revenue.
NVIDIA collaborated on sparse-model research
Sakana and NVIDIA developed sparse data layouts and GPU kernels for language-model training and inference. Their work studies whether sparse activations can produce practical savings on H100 hardware. Sakana publishes four SparseLM checkpoints with the implementation.
Namazu began with three adapted prototypes
The March 2026 alpha series adapted DeepSeek-V3.1-Terminus, Llama-3.1-405B and gpt-oss-120B. Sakana launched the series in Sakana Chat and planned weight releases. Current access to those exact prototypes remains unconfirmed.
Fugu coordinates models from other developers
Fugu routes work among a pool of models. Its standard, Ultra, Max and Cyber options describe the orchestration service. Sakana's API documentation contrasts that system with Namazu, a single model that Sakana adapted. Fugu's service variants therefore do not receive trained-model records here.
Research releases include more than chat models
Practical guidance. The catalog includes published classifiers, maze policies, hypernetworks and voxel generators as well as language models. TinySwallow formats stay with the instruction model. DreamCubed training datasets are excluded from the model count. Demo checkpoints remain with their trained model family.
Each release has its own terms
Practical guidance. Some cards list Apache or MIT while separately describing research-only intent. Both statements are preserved. Several newer repositories provide weights without a clear checkpoint license. Code licensing does not establish the terms for those weights.
FwPKM releases study learned memory
The FwPKM study trains small language models inspired by Qwen3-Next's design. It compares full, linear and sliding-window attention with static or fast-weight memory. These are Sakana research checkpoints, not mirrored Qwen3-Next weights. The fast-weight variants update a sparse memory while reading input.