Sakana AI Moves From Remixing Models to Building Domestic AI From Scratch

Sakana AI will build domestic AI models from scratch, the Tokyo startup said, shifting from adapting existing models to creating originals. Crypto Briefing reported the plan on October 26, 2026. The company raised ¥20 billion, about $135 million, in a November 2025 Series B round at a $2.65 billion post-money valuation.

Sakana AI's Shift From Remixing to Originals

Until now, Sakana AI's approach rested on two techniques. The first is post-training, which refines an existing foundation model for specific tasks and audiences. The second is evolutionary model merging, which combines traits from different models, keeps what works and discards the rest. The company sold that approach on efficiency and cultural alignment, tuning existing models for Japanese language, values and use cases rather than training a giant model from zero.

The new plan puts Sakana closer to peers that build AI frameworks from the ground up, at least for its domestic models. Sakana has pitched itself as a hedge against reliance on foreign models, particularly in finance and defense. The company was founded in Tokyo in 2023 by three former Google researchers: David Ha, Llion Jones and Ren Ito. As of mid-2026, it employed an estimated 157 people.

Sakana's product lineup includes the Namazu family of models tailored to Japanese applications, Fugu, an orchestration platform that coordinates multiple AI agents on a single task, and Marlin, a research agent. Aquarium, a set of enterprise tools, was announced on October 23, 2026.

Sakana AI's Funding and Commercial Products

According to akta.pro company records, Sakana AI launched Sakana Fugu and Fugu Ultra to general availability on June 22, 2026 as a multi-agent orchestration foundation model delivered via an OpenAI-compatible API, with subscription tiers from $20 to $200 per month and pay-as-you-go token pricing. The same records show Sakana Marlin launched on June 15, 2026 as the company's first commercial B2B product, an autonomous "Virtual CSO" deep-research agent powered by AB-MCTS that produces reports of more than 100 pages in up to 8-hour self-governing reasoning loops, with pay-as-you-go and subscription pricing in Japanese yen. Sakana AI raised $214 million in 2024, and its Series B drew backing from Mitsubishi UFJ Financial Group, Khosla Ventures and MPower Partners, among others.

Sakana AI's Recursive Self-Improvement Lab

On October 23, 2026, Sakana AI launched its Recursive Self-Improvement Lab, known as the RSI Lab, focused on research into AI systems that can improve themselves over repeated cycles. The company has framed that work as aligned with Japan's national strategy for sovereign AI, and has built strategic partnerships with major corporations and participated in government-backed initiatives.

Earlier Sakana AI Research Releases

Sakana AI released QD-Rec on October 23, 2026, an open-source recommendation framework that outputs candidates with diverse reasoning arguments rather than just ranked lists, using a contextual bandit and a large-model proposal generator, DoNews reported. On October 10, 2026, MarkTechPost reported that Sakana AI's Multi-Layered Review system caught 73.43% of core-claim errors in a benchmark, at about $0.47 per review. Crypto Briefing reported on October 2, 2026 that MIT and Sakana AI's SIFT framework uses a language model to referee code changes, achieving 35.1% on Polyglot after 30 steps at about $34 in API fees.