Last updated:
ELYZA
ELYZA's Japanese models — including a Japanese-specialised diffusion language model.
Model Facts
Latest model: ELYZA-Diffusion (Dream 7B)
- Provider
- ELYZA
- Released
- Jan 16, 2026
- Weights
- Open
Full Model List
| Model | Released | Status | Note |
|---|---|---|---|
| ELYZA-Diffusion (Dream 7B) | Jan 16, 2026 | current | Japanese diffusion language model, Base and Instruct, Apache 2.0 |
| ELYZA-Thinking-1.0-Qwen-32B | May 1, 2025 | legacy | Qwen2.5-based reasoning model; month approximate |
- Official model page:
- elyza.ai
- Official documentation:
- huggingface.co/elyza
- Other sources:
- ELYZA-LLM-Diffusion (Gigazine)
What it is
ELYZA is a Japanese AI company. Its latest public release (January 16, 2026) is ELYZA-LLM-Diffusion, a Japanese-specialised diffusion language model built on Dream 7B and trained on 62 billion Japanese tokens, with an instruction-tuned variant. Diffusion models generate text quickly at low compute cost. Earlier releases include a Qwen2.5-based 32B reasoning model; both are Apache 2.0.
Where this information comes from
Details come from the January 2026 release coverage. Release dates, context windows and benchmark scores are vendor-reported unless stated, and harnesses differ between labs, so treat the numbers as indicative.
Notable traits
- Apache 2.0, commercial use allowed.
- Diffusion architecture rather than standard autoregression.
- Japanese-first training.