Built for agents from day one: Sakana AI will pretrain its own model series, Sawara
Sakana AI will pretrain its own agent-native model series, Sawara, from scratch. Sawara Small starts training in December, with weights planned for spring 2027 — a small lab betting that research efficiency can beat GPU counts.

Sakana AI, Japan's most closely watched AI startup, has decided to train its own models from scratch. On October 26, it held its first symposium at Hitotsubashi Hall in Tokyo, with more than 400 people in the room; in the official event report published on November 2, chief scientist Takuya Akiba formally introduced a new series called "Sakana Sawara" (a provisional name): a language model series built entirely from scratch, agent-native first and everything else second. Pre-training starts in December, and the weights for the small first model, Sawara Small, are planned for release in spring 2027 — with the qualifier "if everything goes smoothly" attached.

1. From "using other people's models" to "building our own from scratch"
Sakana's shift is first of all a change of direction. The company was founded in 2023 by David Ha, Llion Jones, and Ren Ito, and its best-known work so far has all rested on one assumption: that good models already exist. It merged existing models with evolutionary algorithms, post-trained open models, and orchestrated several models to work together. Its AI Scientist writes papers automatically, its ALE-Agent beat human experts in live competitive programming contests, and AB-MCTS solves extremely hard problems by calling language models over and over — all of it work done on top of the model layer.
Sawara turns that around. Akiba said it plainly at the symposium: with Namazu behind them, they are expanding from post-training into pre-training. Sawara will not be adapted from any existing model. It will be built from scratch — data mix, tokenizer, training objectives, and weights all under Sakana's own control. And it is designed from the start for agentic work: tool-integrated reasoning, coding, and long-horizon tasks. Not a chat model retrofitted with tools, but a model raised to be an agent from its first training step.
The timeline is specific enough that it cannot be dismissed as a slogan: pre-training of Sawara Small starts in December 2026, weights come in spring 2027, and a larger Sawara Large follows. The name continues the company's fish theme — "sakana" means fish in Japanese, and "sawara" (Spanish mackerel) is a fish too. The name is provisional; the direction is serious. A company famous for doing a lot with a little is, for the first time, taking on the most expensive step in the AI business.
The symposium's lineup also says this was not an ordinary launch. Jürgen Schmidhuber, who joined in September as chief scientific advisor, gave the keynote. He co-authored the 1997 LSTM paper and the 2018 "World Models" paper with David Ha, and he is one of the unavoidable names in the history of deep learning. Sakana did not invite him just to lend credibility — almost everything behind Sawara, from world models to recursive self-improvement, can be traced back to papers he wrote more than thirty years ago.
2. The third pillar fills the hardest gap
To place Sawara, start with the three pillars David Ha laid out in his opening talk. He argued that Japan's AI sovereignty rests on three things: pre-training, the ability to develop the technology from scratch; post-training, aligning models with a country's own values; and orchestration, combining its own models with frontier models from the rest of the world to push performance further. Post-training has Sakana Namazu. Orchestration has Sakana Fugu. What was missing was the first pillar: pre-training. Sawara fills it.
Namazu is the best example of how this company thinks. It is an open model post-trained mainly for Japanese users: reducing geopolitical bias, deepening its understanding of Japanese culture, and teaching it ways of communicating and thinking that feel more natural to Japanese users. According to independent reporting, the latest Namazu is built on Moonshot AI's open-weight Kimi K2.6. But post-training has a clear ceiling: the base model sets the upper bound, and no amount of tuning can add what the base model does not contain. Sakana evidently hit that ceiling, which is why it decided to dig one layer deeper.
Fugu is the attempt at the other end: an orchestration model released in June, built to judge when to hand a task to which frontier model and how to combine their answers into one reliable result. Its logic is "rather than building one strongest model, organize the strongest models." That idea has particular bite in an environment where US export controls have tightened access to Anthropic's top models — Fugu is a product that grew out of exactly that constraint.
Put the three pillars together and Sakana's bet becomes clear: Namazu handles "sounding Japanese," Fugu handles "borrowing everyone else's strengths," and Sawara handles "growing something of our own from the root." They are not three parallel lines but a division of labor: pre-training decides how far you can go, post-training decides who you sound like, and orchestration decides what you can deliver today. Sawara is the most expensive, slowest, and hardest-to-reverse of the three — which is why it was announced so carefully, down to the "provisional" label on its name.
3. Akiba's accounting: the model race is not about GPU counts
The most interesting passage in the symposium was not about Sawara itself, but about what Akiba thinks the model race is actually about. Many people, he said, think this competition is about who has more GPUs. That is 100% true — but people also assume GPUs are mainly spent on scaling up: more parameters, more pre-training tokens. That is a misunderstanding. Citing Epoch AI, he noted that at OpenAI, final training runs are estimated to account for only about 10% of R&D compute. The bulk goes into trial and error — in other words, into research itself. Building language models is more of a research competition than most people imagine.
That argument is the key to whether Sawara is feasible at all. If the race were about who owns the most chips, Sakana would not even sit down at the table — independent reporting puts the company at around 20 people. But if the race is about "how much you learn from each experiment," a small team has a chance. Akiba's formula is that progress equals compute multiplied by research efficiency. Compute is public and unwinnable; research efficiency is the hidden variable — who designs experiments that yield the most information. He said plainly at the symposium that his team is the most talent-dense he has ever seen. It sounds like self-praise, but inside the "research efficiency" frame, it is the entire foundation of the Sawara plan.
There is also an organizational footnote worth recording: this summer Sakana set up the RSI Lab, with three teams — language models, world models, and recursive self-improvement — with Akiba leading the language model team. Sawara is not one researcher's side project; it is an institutional commitment by a lab. The three teams map onto the two halves of Sawara: the language model team builds "the model that thinks," the world model team builds "the model that checks," and the recursive self-improvement team connects them into a loop that improves itself. Betting on a direction with a lab's structure, rather than with a single launch event, is a different kind of promise — it means people and budgets will keep flowing in that direction for years.
More striking still is the idea of "fractional recursive self-improvement." Sakana already uses AI to help with its own research; the next step is that some of that AI could be Sawara itself, and the step after that, the next Sawara could be developed with Sawara's help. Akiba called this fractional RSI: not machines designing themselves overnight, but handing over more and more of the research process to the lab's own models, piece by piece. The vision sounds grand, but its first step is very concrete: get the December pre-training run going, and end up with a base model that fully belongs to Sakana.
For onlookers, this argument supplies a standard for judging Sawara. It should not be judged only by parameter counts and leaderboard scores at launch, but by whether Sakana can cash in its "research efficiency" claim. If a team of around 20 people can genuinely catch up with big labs using less compute and smarter experiment design, that is the part of this story that belongs in a textbook. If Sawara Small slips past spring, or lands with a shrug, that will also be an answer: the road of "small team wins on efficiency" is harder than it looks.
4. David Ha's other honest line: base models are "the fastest-depreciating assets"
David Ha said something in his opening talk that does not sound like a CEO talking. Foundation models, he said, are among the fastest-depreciating financial assets in the history of mankind. That takes some nerve from the founder of a company that makes models — but it is exactly why Sakana is moving into pre-training now. If a base model is destined to depreciate fast, then the real value is not in the model itself, but in the capabilities built around it: data, recipes, experimental methods, and the loop that lets models improve themselves. A model you buy will depreciate; the ability to build models will not.
His other core belief is that "intelligence emerges from constraints." Modern AI gets stronger by piling on compute and data, but biological intelligence evolved in scarcity: brains, language, and culture are all proof of doing more with less. He told the story of an experiment from more than ten years ago: when he constrained the size and energy of a bipedal walker in OpenAI Gym, the AI evolved a small, insect-like walker; with all constraints removed, it built a giant brainless robot that simply fell over toward the exit. Constraints are not obstacles. They are what give AI a brain in the first place.
Put the two lines together and you have the complete logic of Sawara: limited resources are not a disadvantage, they are a method. Japan does not have America's compute or China's scale, but it has robots, manufacturing, and a tradition of doing things to the extreme within constraints. Schmidhuber echoed the point in his keynote: in 2002, more than half of the world's robots came from Japan, while today "the only AI that works really well is AI behind the screen" — no robot comes close to what a plumber or an electrician can do, or what a little boy can do with a football. Screen-bound AI already has owners. Physical AI does not yet, and robot-oriented countries like Japan and Germany have a chance to lead there.
Vision aside, reality still bites. Schmidhuber himself admitted that robot learning has two gaps left: speed and safety. Driving robots directly with large language models is slow and unreliable — in Sakana's own experiments, 18% to 28% of episodes do not follow instructions exactly. Their answer is the world model: run the action through a mental simulation first, judge feasibility, predict failures, and only try the promising plans on real hardware, then feed the results back to improve the model. That "think before acting" architecture is exactly what the world-model half of Sawara is for: the language model does the thinking, the world model does the checking, and together they make an agent that can act without causing trouble.
5. What can be said now, and what cannot
At this point, the actual status of Sawara has to be stated plainly: today it is a plan, not a product. The official report contains no model size, no data volume, no compute budget, no license, and not even a clear statement of what "releasing the weights" means in practice — downloadable weights, a hosted API, or both. Only three things are certain: pre-training starts in December, Sawara Small's weights come in spring 2027, and a larger Sawara Large follows later. Nikkei's reporting adds two more claims — the larger model is expected to reach several hundred billion parameters, and the effort is positioned around defense and cybersecurity — but those come from Nikkei's reporting, not from Sakana's official report, and the sourcing should not be blurred.
This style of announcement — direction only, no details — is common in the AI industry, but it should not be confused with a launch. Sawara has no benchmark scores, no demo, and no model card. Its only verifiable deliverable is the wait until spring 2027. Sakana itself left an escape hatch in the announcement: "if everything goes smoothly." Those five words are the entire disclaimer. Whether the December pre-training starts on time, and whether the weights arrive in spring as promised, is the first hard test of this company's execution.
But "no details" should not be read as "no substance." Sakana's record over the past three years is the credit behind this plan: the AI Scientist was published in Nature this year, and a paper written by the third-generation Agentic Scientist was accepted as a main conference paper at ALIFE 2026; ALE-Agent beat human experts in live competitive programming; Fugu's orchestration idea has been publicly validated since its June release. A lab that keeps producing real work announcing a "see you next spring" plan carries a different weight from a company that only draws roadmaps.
6. What it means for vibe coders
For people who write code with AI every day, the most interesting thing about Sawara is not its parameter count, but its design goal: agent-native. A model optimized from the pre-training stage for tool use, coding, and long-horizon tasks will feel different from a chat model retrofitted into an agent — potentially steadier in when it calls tools, more coherent across multi-step tasks, and better at recovering from errors. Sakana is betting that difference is worth starting from zero: rather than patching up someone else's base model, grow the model as an agent from day one.
If the bet pays off, the benefit is not Sakana's alone. Sawara Small's weights are planned for public release, which means independent developers and small teams could gain a new option next spring: an openly available weight set tuned specifically for agent work — usable as a local coding assistant, fine-tunable, and open to study. Sakana has a track record of publishing its methods — the AI Scientist's code and papers are public, and Fugu has a technical report. Whether Sawara continues that tradition, and under what license, is the second question to watch next spring, right after "does it actually ship on time."
Sawara's outcome also matters beyond itself, because it answers a question about the industry: is pre-training now a game only giants can play? If a lab of around 20 people can genuinely build a usable agent-native model on research efficiency, then the door to pre-training is not closed, and smaller teams still have a ticket in. If it slips or lands flat, that is an answer too: the scale bar for pre-training has risen past the point where being clever can compensate. Either way, next spring's release will be worth recording for more than the release itself.
Three things are worth watching next, and all three are concrete. First, whether pre-training starts on schedule in December — the first step from paper to reality. Second, whether Sawara Small's weights arrive on time in spring, under what license, and with what benchmark scores — that decides whether it is a genuine open asset or a limited showcase. Third, whether a timetable and scale for Sawara Large follow after Small ships — that is the real test of whether the "from scratch" road can reach the frontier.
To sum up: Sawara is not a model launch. It is a small lab's statement about how the AI industry should be organized — that the good models of the agent era should not be retrofitted chat models, but built for getting work done from the first day of pre-training. Whether that judgment is right will only have its first piece of evidence next spring. But the judgment itself is already worth writing down today.
Comments (0)
Sources
Related articles
On 20 November 2026, Thinking Machines Lab released Inkling 2 Mini: a 26.4B-parameter looped transformer that runs its middle 12 layers three times per token, fits in about 20GB, and runs offline on a laptop. It leads open-weights models of its size on DeepSWE and SWE-bench Verified — with the caveats printed in the same announcement.
On 17 November 2026, Microsoft launched MAI Code 2 and Code 2 Flash: the larger model has 77 billion active parameters and a one-million-token context, averaging 63.2% across four benchmarks — close to GPT-6.1 Sol and Claude Opus 5.5 — at $0.309 per task. Read the three caveats first: promotional pricing, the lowest score of the three, and no long-horizon work.
On 17 November 2026, Perplexity launched the Orchestrator API: one interface reaching OpenAI, Anthropic, Google, xAI and open-weight models such as GLM, Kimi and DeepSeek, billed at the answering model's published rate with no platform fee. When one provider fails, the same model retries on another deployment, and preview routing assigns requests to models by task.