Sakana AI Bets on Sawara to Build Agent-Native Models From Scratch

Tokyo-based Sakana AI used its first public symposium to announce that it will pre-train its own foundation models. Previous releases, including the Japanese-tuned Namazu, adapted existing open-weight models. The new Sakana Sawara series will give the company control over the data mix, tokenizer, training objectives, and model weights from the start.

The Tokyo event drew more than 400 attendees and featured Jürgen Schmidhuber, who became Sakana’s chief scientific adviser in September. Schmidhuber joined co-founder and CEO David Ha, his co-author on the 2018 World Models paper, for a keynote and fireside discussion. Their talks placed Sawara within Sakana’s broader research on AI-generated experiments, recursive model development, and robotics.

An agent starts at pre-training

Chief scientist Takuya Akiba said the company plans to release Sawara Small in spring 2027, subject to development progress. A larger Sawara Large model is planned later, although Sakana has not announced its size, schedule, or compute budget. The company has promised public weights but has not specified the license or accompanying training artifacts.

“Agent-native” has no settled technical definition. Sakana uses the term for a model whose pre-training data and objectives incorporate tool use, coding, planning, and multi-step task completion. An agent can choose actions, call external software, inspect results, and continue working toward a goal across multiple steps.

Sawara will also support the company’s recursive self-improvement research at its RSI Lab. Akiba described a limited form he called “fractional RSI,” in which one model helps researchers design data, evaluations, or training procedures for its successor. Sakana has not disclosed which parts of Sawara’s development will be automated or how human review will govern those changes.

Control of the training recipe

Pre-training creates a model’s initial capabilities by exposing it to large datasets and adjusting its weights to predict or generate content. Post-training then specializes those weights through instruction tuning, preference optimization, reinforcement learning, or other targeted methods. Decisions made during pre-training shape the model’s language coverage, coding ability, token efficiency, and capacity to learn later behaviors.

Akiba argued that model development depends on research iteration as well as the final training run. He cited an Epoch AI estimate that final runs account for about 10% of OpenAI’s research and development compute, with experiments consuming the remainder. He summarized Sakana’s approach as compute multiplied by research efficiency, suggesting that a smaller team can compete by running better experiments. Sakana has not published the estimate behind its own compute requirements.

Third-party open models often provide weights without complete access to their original corpora, filtering rules, tokenizer decisions, intermediate checkpoints, or training recipes. Owning those components allows Sakana to vary them, measure the effects, and let future models propose changes. That access is particularly relevant to recursive self-improvement experiments, which require reproducible control over the training process.

Robots get a rehearsal layer

Sebastian Risi described a robotics architecture that places a world model between a vision-language model and physical hardware. The vision-language model interprets images and instructions, then proposes actions. The world model acts as an internal simulator, predicting the likely consequences before the robot moves.

The proposed control loop

1. A compact vision-language model quickly generates several candidate action plans.
2. A world model simulates the next steps, estimates feasibility, and predicts likely failures.
3. The controller sends approved plans to the physical robot.
4. Real-world outcomes update the vision-language model after passing safeguards intended to limit reward hacking, where a system exploits an evaluation rule without completing the intended task.

Risi said that 18% to 28% of episodes labeled successful in evaluations of current language-model-driven robot policies still failed to follow the instruction. He also reported an early experiment with a fine-tuned π0.5 policy and a simulated Franka arm, conducted with Japan’s AI Robot Association, in which the loop improved a drawer-opening task from failure to success. Sakana has not published sample counts, baseline variance, evaluation code, or real-hardware results, so the demonstration remains preliminary.

Schmidhuber’s 1990 loop returns

Schmidhuber connected the robotics work to a controller and world-model architecture he proposed in 1990. The controller selects actions while the world model forecasts their effects, allowing the system to evaluate behavior internally before interacting with its environment. Modern neural networks and faster hardware make the same structure practical at a larger scale.

His keynote characterized current generative AI as part of a broader research program that includes artificial scientists capable of designing experiments and robots capable of learning to use physical tools. Sakana’s AI Scientist work addresses the first area by automating parts of the research cycle; its world-model program targets the second.

Schmidhuber also argued that the price of a fixed amount of compute falls by roughly a factor of 10 every five years, giving smaller laboratories access to capabilities that once required larger budgets. He described progress in physical AI as slower, estimating that Japanese walking robots from 30 years ago trail current systems by a factor of three to five. The symposium did not specify the metric behind that comparison, but the claim reflects the longer testing cycles and hardware constraints of robotics.

What Sawara could add

• A Japan-based training stack: Sawara would add an independently pre-trained model to a field dominated by US and Chinese laboratories.
• Public weights: Sakana’s release commitment could support inspection and adaptation, depending on the final license and available artifacts.
• A smaller research model: Sawara Small could reduce the cost of studying agent-focused pre-training if Sakana releases sufficient technical detail.
• A shared research substrate: Sakana plans to use the series across tool-using agents, AI-generated research, recursive training experiments, and world models.

Specifications still undisclosed

• Parameter counts, architecture, context length, and memory design
• Training languages, corpus sources, filtering methods, and data governance
• Tokenizer design, compute budget, hardware, and training duration
• Tool protocols, agent scaffolding, and supported execution environments
• Evaluation suites, baselines, safety tests, and long-horizon task metrics
• Weight licenses, checkpoints, training code, and dataset availability

Sawara’s technical case will depend on measurable tool reliability, long-horizon task completion, training efficiency, and reproducibility when Sawara Small arrives. The robotics program requires broader simulated tasks and controlled physical trials. Those results will show whether Sakana’s agent-focused pre-training and controller-plus-world-model architecture produce gains that other researchers can verify.