ELYZA
★ 3/5 · Open Source Model

ELYZA, a University of Tokyo spin-out now owned by KDDI, releases Japanese-tuned open-weight models on Hugging Face — including Llama-3-ELYZA-JP-8B and the ELYZA-Thinking and ELYZA-Shortcut reasoning models — while its commercial business is Japanese enterprise consulting and the ELYZA Works internal-AI-app tool. The models are free to download; there is no ELYZA-hosted API.
Pros
- Eighteen public model repositories covering the most useful sizes for Japanese work — Llama-3-ELYZA-JP-8B at 8B (about 6.5k downloads a month, with AWQ and GGUF builds published alongside it so it runs on a single consumer GPU or under llama.cpp), the 32B ELYZA-Thinking and ELYZA-Shortcut reasoning models, and ELYZA-Diffusion-1.0-Dream-7B from January 2026, a Japanese diffusion language model that generates from noise rather than left to right. Model cards are in English with runnable transformers and vLLM snippets, and the company is visibly still shipping: an Agentic RL write-up in March 2026, a Japanese medical LLM built for the Cabinet Office SIP programme, and named production deployments at Mitsui Sumitomo Card and Altius Link.
Cons
- There is no ELYZA API, console or hosted endpoint of any kind — the open weights are the entire self-serve surface, and everything commercial runs through a Japanese enterprise sales conversation. elyza.ai is Japanese-only with no English version, no published price list for either the consulting business or ELYZA Works (resold to Japanese corporates through KDDI), and no English support path. The flagship Llama-3-ELYZA-JP-8B sits on the Meta Llama 3 Community Licence rather than a permissive one, dates from June 2024 and publishes no benchmark table on its card, and the newer reasoning and diffusion models draw a fraction of that traffic — the 32B Thinking model shows around 125 downloads a month.
Cost
The open weights on Hugging Face are free to download, subject to the licence attached to each repository — Meta Llama 3 Community Licence for the Llama-3-ELYZA-JP line, with the Qwen-derived and diffusion models carrying their own terms — and your only cost is the hardware you run them on. ELYZA publishes no pricing at all for its commercial work: the AI Research & Solutions consulting business and the ELYZA Works product are both quoted per engagement through Japanese sales, with ELYZA Works sold via KDDI.
Verdict
Best for developers who specifically need a small, well-tuned Japanese model to run locally and are happy to work from a Hugging Face card; treat ELYZA the company as unavailable unless you are a Japanese enterprise able to run a procurement process in Japanese.
Visit ELYZA