Stability AI launches its first Japanese open model

Japanese StableLM Alpha is a new 7B general-purpose language model designed specifically for Japanese text generation and instruction following. If you need a capable Japanese model that you can run locally or fine-tune yourself, this is arguably the strongest open release available right now.

The release includes two distinct flavors: Japanese StableLM Base Alpha 7B for text generation, and Japanese StableLM Instruct Alpha 7B, which has been tuned with supervised fine-tuning to follow conversational prompts.

What is under the hood

Stability AI built this on an extension of EleutherAI's GPT-NeoX software, adopting modern architecture tricks like SwiGLU and xPos. Across multiple training epochs, the model chewed through 750 billion tokens.

The training corpus consists predominantly of Japanese and English web text, with source code making up the remaining 2 percent. Stability AI Japan sourced data not just from standard open pools, but also via custom datasets created alongside the EleutherAI Polyglot team and the Japanese open-source community.

On the benchmark front, the Instruct version delivers. Evaluated with EleutherAI's lm-evaluation-harness across eight core tasks covering question answering, sentence classification, and summarization, it scored an average of 54.71. That comfortably beats four other established Japanese open models tested on the same setup.

Licensing and availability

Both checkpoints are up on Hugging Face Hub right now, ready for inference and experimentation.

Pay attention to the licenses before shipping code. The Base Alpha 7B model is released under Apache 2.0, meaning you can drop it straight into commercial software and products.

However, the top-performing Instruct Alpha 7B model is restricted strictly to research purposes. If you want a commercial chat product, you will have to fine-tune the Base model on your own instruction datasets.

Why it matters

Building non-English models that actually understand cultural context and grammar has always been an uphill battle for open-source AI. Having a reliable 7B foundation dedicated to Japanese with commercial-friendly Apache 2.0 terms is a massive win for indie builders and enterprises looking to avoid closed API lock-in.

The research-only tag on the Instruct checkpoint is an annoyance, but having the Base model ready to fine-tune makes it well worth spinning up a GPU instance today.