Rakuten AI
Rakuten publishes a family of Japanese-optimized open-weight LLMs on Hugging Face under Apache 2.0, from the 1.5B-parameter Rakuten AI 2.0 mini up to the 671B-parameter Rakuten AI 3.0 mixture-of-experts model. There is no Rakuten-hosted API — you download the weights and serve them yourself.
Pros
- Apache 2.0 across the whole line, about as permissive as open weights get from a company this size — commercial use, fine-tuning and redistribution all allowed with no separate agreement. The lineup spans real deployment sizes: Rakuten AI 2.0 mini at 1.5B for on-device and edge work, Rakuten AI 2.0 8x7B at 47B total and 13B active (scoring 7.08 on Japanese MT-Bench), and Rakuten AI 3.0 at 671B total with 37B active and a 128K context window, trained partly under METI and NEDO's GENIAC programme and released in March 2026. Model cards, benchmark tables and SGLang/Docker serving instructions are all in English, base and instruct checkpoints ship together, and the community has already produced GGUF and quantized builds.
Cons
- Weights only — there is no Rakuten API, console, playground or hosted endpoint, so the cost of running these is entirely your own GPU bill, and the 671B flagship realistically needs an eight-way tensor-parallel node before it will load at all. Adoption is thin: the Rakuten AI 3.0 repository shows roughly 121 downloads in the last month against 6.5k for smaller Japanese models elsewhere, so you are early and largely on your own for tooling, and Rakuten explicitly warns the models can produce biased or unsafe output with no hosted guardrail layer to fall back on. English and multilingual quality are secondary to Japanese throughout the family.
Cost
Free. Every Rakuten AI model is released under Apache 2.0 at no charge from the official Rakuten Group Hugging Face repository, with no licence fee, API pricing or usage tier published — your only cost is the compute you provide to run them, which for the 671B Rakuten AI 3.0 means multi-GPU serving infrastructure.
Verdict
Best for Japanese teams and researchers who want a permissively licensed Japanese model they can host, fine-tune and ship inside their own infrastructure without a vendor contract; not the choice if you want to call an endpoint tomorrow, because nobody at Rakuten will host it for you.
Visit Rakuten AI