Sakana AI Builds Sawara, Its First Foundation Model From Scratch

Sakana AI plans to build a large language model from randomly initialized weights, expanding its work beyond model merging, post-training, and agent orchestration. The Tokyo lab’s Sawara announcement says pretraining for the first checkpoint will begin next month. Public weights for Sawara Small are targeted for spring 2027, with Sawara Large planned to follow. Random initialization means the training run begins without inherited weights, the numerical parameters that encode another model’s learned behavior.

Owning the base model

Sakana built its earlier work around two layers. Its agent portfolio includes ALE-Agent for combinatorial optimization and AB-MCTS, which explores candidate reasoning paths while a model is answering. At the model layer, Sakana Namazu adapts foreign open-weight models for Japanese use, while Sakana Fugu coordinates several models to improve performance.

Fugu’s capabilities depend on the models available from outside labs, including releases from the United States and China. Sawara gives Sakana control over the base architecture, training data, optimization process, release schedule, and later specialization. It also lets the lab test whether its agent research can shape a model throughout development.

Small before Large

Chief Scientist Takuya Akiba will lead a small, agent-assisted team. Sakana is considering the construction of models with tens of billions of parameters or more.

Cheaper training changes the calculus

• Lower development costs: Improvements in hardware, training software, and published techniques have reduced the resources required to produce capable models.
• Stronger open-weight models: The narrowing gap with leading proprietary systems suggests that useful training methods are spreading beyond the largest laboratories.
• Task-specific reinforcement learning: Rewards based on task outcomes can tune a pretrained model for coding, tool use, and other agent workflows.

Sakana also cites external analyses indicating that experimentation can consume more compute than the final training run. The company argues that rapid iteration, careful experiment selection, and concentrated research talent can offset some of the advantage held by labs with larger clusters. Compute remains a constraint, particularly for Sawara Large.

What “agent-native” means

Sakana uses “agent-native” to describe a model designed for tool calls, coding, and long-horizon tasks that require many coordinated steps. Tool use includes selecting and calling external software or APIs. Long-horizon performance depends on maintaining context, checking intermediate results, and recovering from errors across an extended sequence of actions.

The company expects agent-oriented training and reinforcement learning to improve those capabilities throughout the model’s development. It has yet to specify whether that approach will rely on specialized pretraining data, architectural changes, new training objectives, post-training methods, or a combination of them.

AI agents will also help plan pretraining experiments. That process extends Sakana’s AI Scientist work and its RSI Lab, which studies systems that can improve AI research workflows. The company has not defined which experimental decisions agents will make autonomously or what review process human researchers will use.

The missing specifications

• Compute: No training budget, accelerator count, or estimated training time has been disclosed.
• Model size: Sakana has given a broad target of hundreds of billions of parameters without exact counts for either model.
• Training data: The corpus, language mix, token count, filtering process, and data licenses remain unspecified.
• Architecture: The company has not published the context window, tokenizer, model architecture, or inference requirements.
• Evaluation: No benchmark suite or baseline models have been named for agent, coding, Japanese-language, or safety testing.
• Release terms: Sawara Small’s weights are planned for public release, but the license and availability of training code or data have not been announced.

Japan’s crowded model race

Several Japanese companies began pretraining domestic foundation models before Sakana entered the field. Sakana’s strategy depends on research efficiency, agent-assisted experimentation, and task specialization compensating for its later start and smaller compute base. A successful Sawara release would also reduce the company’s dependence on foreign model roadmaps.

Sakana has said it may pursue an initial public offering as early as fall 2027 and is considering simultaneous listings in Japan and the United States. The company plans to direct proceeds toward proprietary model development and hiring, which would support the more capital-intensive Sawara Large program.

Four tests for Sawara Small

1. Release usability: The license, hardware requirements, tokenizer, and inference tooling will determine whether developers can adopt or fine-tune the model.
2. Reproducibility: Documentation covering data sources, training stages, hyperparameters, and evaluation methods would let researchers inspect and extend the work.
3. Agent performance: Coding, tool-use, and long-horizon evaluations should compare Sawara Small with similarly sized open-weight baselines.
4. Training efficiency: Disclosed compute, staffing, and experiment counts would provide evidence for Sakana’s claim that a small, agent-assisted team can compete through faster research iteration.