Sakana AI will train Sawara, its first LLM built from scratch

Felix Laurent 13 Nov 2026 4 min read

Models

Sakana Sawara is the first language model that Tokyo startup Sakana AI will train from scratch. The company plans to start pretraining the first version, Sawara Small, in December 2026. If the work goes smoothly, it will release the weights in spring 2027.

Sakana first presented the model at its symposium on October 26. On November 13, it explained the decision in written answers to Japanese outlet ITmedia AI+. "The time has come," the company said of its move into pretraining.

The shift is a notable change of course. Sakana built its name by reusing other companies' models rather than training huge ones of its own.

What Sakana Sawara is

Sawara is a family of models built for AI agents. Sakana calls it an "agent-native" language model, designed from the start for tool use, coding and long-running tasks. All names are tentative for now.

Chief Scientist Takuya Akiba presented two sizes at the symposium. Sawara Small comes first and is meant as a pilot for later models, though it can also be used on its own. Sawara Large would be much bigger. ITmedia reports that Sakana is considering models with several hundred billion parameters or more.

Sakana has not disclosed the parameter count of Sawara Small, the size of its training data, or the hardware it will use. It also has not said what license the weights will carry. The company only says a small, elite team will build it.

The model will serve two purposes. It will become the base for Sakana's products and services, and a base for its research into AI that improves itself. To see why that matters, it helps to look at what Sakana did before.

Why Sakana AI is moving into pretraining now

Until now, Sakana worked mostly on top of other companies' models. Its first research merged existing open models into new ones. This year it released Sakana Namazu, open foreign models post-trained for Japanese users. Its Fugu system combines frontier models from other labs with Sakana's own.

Sakana says three changes made pretraining practical for a company of its size. Open-weight models have become good enough to use in real work. Training data and recipes are easier to get, and development costs have fallen. Agentic reinforcement learning can now tune a model end to end for one company's tools and tasks.

Owning the whole model is the main argument. "Being able to understand the whole model and training recipe" opens up broader research, Sakana told ITmedia. It added that every step of development becomes transparent to the company.

Sakana also argues that model building is less about raw compute than people think. It points to an Epoch AI estimate that only about 10% of OpenAI's R&D compute went into final training runs. In Sakana's view, the rest goes into experiments, so research talent matters as much as GPUs.

That view ties directly into the lab Sakana opened this summer.

How Sawara fits Sakana's self-improving AI research

In June 2026 Sakana launched RSI Lab, short for recursive self-improvement. The lab studies how AI can help develop the next generation of AI. Akiba leads its language model team, alongside teams for world models and RSI research.

Sawara is meant to be a test bed for that idea. Sakana describes a "partial RSI" plan, in which later versions of Sawara could be developed partly with help from Sawara itself. A model the company knows down to every training step is easier to study that way than someone else's open weights.

The lab builds on Sakana's earlier research. Its AI Scientist project, first shown in 2024, automates parts of machine learning research and was published in Nature in March. Its ALE-Agent took first place against human experts in a live programming contest.

Sakana also works on defense and security projects. It declined to discuss specific requirements there, but said that a model independent of others gives it full control. That point also matters for Japan's wider AI plans.

Where Sawara fits among Japan's own models

Sakana frames Sawara as one of three pillars of AI sovereignty, alongside post-training with Namazu and orchestration with Fugu. Several companies in the US, China and Japan already pretrain their own models. Sakana admits it is following them.

In Japan, Preferred Networks builds its PLaMo models, SoftBank's SB Intuitions trains Sarashina, and the national LLM-jp project releases open models. Sakana's bet is that an agent-focused design and research efficiency can make up for a late start.

The move also comes as the company prepares for a possible public listing. Nikkei reported this month that Sakana AI is preparing an IPO for as early as autumn 2027. A home-grown model would give investors a clearer story than a business built on other labs' models.

Can Sakana AI build a competitive model from scratch

It can build a useful one, but probably not a frontier one at first. Sawara Small is a pilot, and Sakana has not shared its size or budget. A model trained by a small team will not match the largest US and Chinese systems on general benchmarks.

That may not be the goal. Sakana wants a model it fully controls, tuned for agents and Japanese customers, and open to its own self-improvement experiments. Sawara Small has to ship on time next spring first. After that, Sawara Large will show whether research efficiency can stand in for a giant compute budget.

FAQ

Here are short answers to common questions about Sawara.

When will Sakana Sawara be released?

Sakana plans to start pretraining Sawara Small in December 2026. If training goes smoothly, it will publish the weights in spring 2027. Sawara Large has no date yet.