Japan's LLM-jp-4-VL 9B Beats InternVL on Japanese While Staying Truly Open Source

Japan's national LLM initiative has shipped LLM-jp-4-VL 9B, a bilingual vision-language model pairing a reasoning-tuned Japanese LLM with a SigLIP 2 vision encoder. Released under Apache 2.0, the model generates an explicit chain of thought before answering and was trained exclusively on data with clean commercial licensing, making it one of the few VLMs with genuine open-source compliance and strong Japanese support.

What's under the hood

The architecture draws from InternVL3.0 and combines three components:

• LLM: llm-jp/llm-jp-4-8b-thinking (8.6B)
• Vision Encoder: google/siglip2-so400m-patch16-512 (0.4B)
• Projector: 2-layer MLP

Because the base LLM is already a reasoning model, each turn produces two outputs: an analysis message containing the chain-of-thought reasoning, followed by a final answer. The chat template extends OpenAI's Harmony format with <|image_start|> and <|image_end|> tokens wrapped around image patches inserted into the user message.

The licensing problem and how they solved it

Much open VLM training data is contaminated with outputs from closed proprietary models, which bars downstream commercial use. To address this, the team built RefinedVision, a cleaned rework of HuggingFaceM4's FineVision. Of the 185 subsets in FineVision, they dropped 61 flagged for licensing concerns or poor quality. For 58 of the remaining 124 subsets, whose original QA pairs were generated by GPT-4o or other proprietary models, they discarded the QA entirely and regenerated it from scratch using Qwen3.5 397B-A17B with per-category prompts, keeping the original images throughout.