The global AI race just got a powerful new contender — and this time, it's coming straight out of Japan.
On April 3, 2026, the National Institute of Informatics (NII), one of Japan's most prestigious publicly funded research institutions, officially released two new large language models under an open-source license: LLM-jp-4 8B and LLM-jp-4 32B-A3B. These models were trained on approximately 12 trillion tokens of high-quality corpus data — a massive leap in scale and quality compared to previous Japanese-developed AI models.
What makes this release truly headline-worthy? On several key benchmarks, LLM-jp-4 has managed to outperform both GPT-4o (OpenAI's flagship model) and Qwen3-8B (Alibaba's competitive Chinese LLM). That's not a small claim. For the AI community in Southeast Asia and beyond, this announcement deserves serious attention.
What Is LLM-jp-4, and Why Should You Care?
LLM-jp is a large language model project led by NII — a national inter-university research institution in Japan that operates under the country's Ministry of Education, Culture, Sports, Science and Technology. In simpler terms, this is a government-backed, academically developed AI — not a corporate product built primarily for profit.
The latest iteration, LLM-jp-4, comes in two variants:
- LLM-jp-4 8B — An 8-billion parameter model, compact enough to run on consumer-grade hardware with the right setup.
- LLM-jp-4 32B-A3B — A larger, more powerful 32-billion parameter model using a Mixture of Experts (MoE) architecture, activating only 3 billion parameters at a time for greater efficiency.
Both models were trained on roughly 12 trillion tokens, which include carefully curated Japanese-language data alongside multilingual sources. The emphasis on quality corpus — not just raw data volume — is what sets this project apart from many rushed AI releases.
---
Benchmark Results: How Does It Stack Up?
Here's where things get exciting. According to NII's official release, LLM-jp-4 achieves performance that surpasses GPT-4o and Qwen3-8B on certain benchmarks, particularly those involving Japanese-language understanding, reasoning, and knowledge tasks.
To put that in perspective:
- GPT-4o is OpenAI's most advanced publicly available model as of 2025-2026, widely regarded as a gold standard for general-purpose AI.
- Qwen3-8B is Alibaba's latest open-source model, which has been dominating leaderboards in Asia with its strong multilingual capabilities.