loading
loading
Sakana AI Plans Sawara, a Japanese Foundation Model Built From Scratch Sakana AI unveils Sawara, a from-scratch Japanese foundation model, alongside a three-pillar strategy spanning pre-training, post-training, and orchestration. - Sakana AI announced Sakana Sawara, a from-scratch Japanese foundation model targeting hundreds of billions of parameters. - Development starts in November, with a small model shipping in spring 2027 and a larger model to follow. - CEO David Ha laid out three pillars: pre-training (Sawara), post-training (Namazu), and orchestration (Fugu). - Chief Scientist Takuya Akiba will lead a small, agent-assisted team tackling the pre-training work. - Target markets include Japanese defense, cyber defense, and financial services, reducing reliance on US or Chinese APIs. - Jürgen Schmidhuber delivered the keynote as newly appointed Chief Scientific Advisor at the Tokyo symposium. Sakana AI plans to train a Japanese foundation model from randomly initialized weights under a new project tentatively named Sakana Sawara. CEO David Ha announced the work at a symposium at Hitotsubashi Hall in Tokyo, extending the lab beyond the model-merging, post-training, and orchestration systems that built its profile. Foundation models learn general patterns from large datasets and can later be adapted for specific products. According to Nikkei, Sakana expects to begin Sawara development in November 2026, release a smaller model in spring 2027, and pursue larger versions with hundreds of billions of parameters. Why Sakana wants the base layer Sakana became known for extracting more capability from existing models. Projects including the AI Scientist, evolutionary model merging, and the MASS self-improvement work with UC Berkeley operate above pretrained models through agents, routing, or automated experimentation. Chief scientist Takuya Akiba gave three reasons for adding pre-training to that work: - Open-weight models have narrowed the performance gap with leading proprietary systems. - Training costs have fallen enough for smaller labs to consider serious base-model programs. - Reinforcement learning can specialize a base model for particular tasks, tools, and operating environments. Pre-training still requires substantial computing capacity because teams must process vast datasets over repeated experimental runs. Akiba argued that experiment selection has therefore become a central constraint: researchers must choose promising ideas, sequence tests carefully, and stop weak approaches before they consume large GPU allocations. Sakana plans to use AI agents to propose experiments, launch training runs, evaluate results, and help researchers decide what to test next. That approach follows the company’s broader efficiency thesis, which emphasizes research coordination and model combinations. Three layers, three product lines Ha organized Sakana’s work around three layers of the AI stack: pre-training creates a base model, post-training adapts an existing base, and orchestration coordinates models at inference time. Namazu is already in production. Sakana says the latest generation uses Moonshot AI’s open-weight Kimi K2.6 as its base and applies further training for Japanese users and organizations. Fugu handles inference across a pool of models. It decomposes a request, routes each part, and combines the outputs. Sakana offers Fugu Ultra, Fugu Max, and Fugu Cyber, and has reported an 86.9% success rate on CyberGym, a benchmark for cybersecurity agents. Sovereignty reaches the weights Sakana is positioning Sawara for defense, cyber defense, and financial services, where data residency, operational control, and supplier risk shape procurement. Japanese institutions in those sectors may face restrictions on sending sensitive material to foreign-hosted APIs. Starting from randomly initialized weights would give Sakana control over the base model’s data mixture, tokenizer, training objectives, and subsequent alignment process. Namazu inherits its initial weights from Kimi K2.6; Sawara would give the company its own base checkpoint. A full sovereignty assessment will also require details about training-data provenance, chip and cloud suppliers, software dependencies, licensing, security controls, and deployment options. Domestic control over weights addresses one part of that assessment. A staged run toward 2027 Nikkei reported a three-stage schedule that moves from initial research to progressively larger training runs: - Begin development in November 2026. - Release a smaller Sawara model in spring 2027. - Train a larger version targeting hundreds of billions of parameters. The staged plan would let Sakana test its data pipeline, architecture, evaluation suite, and scaling behavior before committing more GPUs to a larger run. Parameter count provides a rough measure of model size; capability will also depend on training data, optimization, architecture, and post-training. Sakana has described a small, agent-assisted research team as an advantage for rapid experimentation. The approach will test whether a lab of roughly 20 people can coordinate a model program at this scale while managing the compute and engineering demands of pre-training. What developers should watch - Access: Sakana has not said whether Sawara will ship as downloadable weights, a hosted API, an on-premises product, or some combination. - License: Commercial terms, redistribution rights, and fine-tuning permissions remain undisclosed. - Architecture: The company has not provided the model family, parameter count for the smaller release, context window, tokenizer, or hardware requirements. - Evaluation: Developers will need Japanese-language, coding, reasoning, security, and domain-specific benchmarks conducted under reproducible settings. - Operations: Inference cost, latency, quantization support, tool-calling interfaces, and deployment requirements will determine practical adoption. Current Sakana integrations center on Namazu and Fugu, and Sawara remains a 2027 research and product target. The project gives the lab control over the deepest layer of its stack and could provide regulated Japanese organizations with a locally controlled base model. Its practical value will depend on measured performance, operating cost, security, licensing, and release terms.
document fetched 2026-10-27 02:55