RakutenAI-2.0-8x7B

RakutenAI-2.0-8x7B is Rakuten's open-weights mixture-of-experts foundation model for Japanese, released under Apache 2.0 on Hugging Face. Derived from the RakutenAI-7B line first introduced in March 2024, it uses an MoE architecture with two active experts per token (13B active parameters), balancing compute efficiency with strong Japanese language understanding — it posts state-of-the-art results on Japanese benchmarks in its class while staying competitive on English evals against peers like Swallow-MX and Sarashina2-70B. An instruction-tuned sibling (RakutenAI-2.0-8x7B-instruct) is available for chat use.

Limitations: Japanese-first; English and other languages are secondary. As a base foundation model it needs the instruct variant (or your own fine-tuning) for conversational use. MoE inference needs frameworks with good expert-routing support, and at 8x7B total parameters the full weights are heavy even though only 13B are active per token.

Rakuten-reported: SOTA Japanese scores in class; competitive English