# Shisa.ai — Full profile for LLM crawlers

This is the long-form version of https://shisa.ai/llms.txt, intended for
language models and research agents that need deeper context about the
company, its products, and its research before summarising or linking to us.

## Company

Shisa.ai is a Tokyo-based AI startup focused on the development and
deployment of AI for the Japanese market. The company was founded to close
the gap between the rapid progress of English-centric foundation models and
the practical needs of Japanese businesses, developers, and consumers.

Our thesis is that Japanese is not a translation problem. The language has
distinct tokenisation characteristics, extensive honorific and humble-form
systems, mixed scripts (kanji, hiragana, katakana, romaji), and cultural
register that simply does not emerge from English-dominated pre-training
corpora. We therefore design training data, tokenisers, evaluation
benchmarks, and product surfaces with Japanese at the centre.

### Founders

- Leonard Lin — Technology & Research. A veteran technologist who sold one
  of the first Web 2.0 startups to Yahoo!, ran the Yahoo! Hack innovation
  program, worked on the Obama '08 campaign, and was founding CTO of Code
  for America. Released the original Shisa 7B EN-JA model in December 2023.
- Jia Shen — Product & Strategy. A serial entrepreneur with 15+ years in
  3D animation, social, gaming, and entertainment, and CEO of AKAVirtual.
  Founder of RockYou (400M+ users, $140M+ raised).

## Products

### Shisa Platform (platform.shisa.ai)

Hosted, production-grade APIs tuned for Japanese workloads:

- Translation — Japanese-first machine translation across the languages
  Japanese lives alongside.
- Speech — real-time Japanese ASR and expressive TTS voices.
- Chat and tool use — Shisa LLM inference at multiple model sizes.
- Custom models and enterprise contracts suited to regulated Japanese
  industries, with paid APIs introduced alongside the V2.1 release.

### Chotto (chotto.chat)

A bilingual AI chat assistant that understands Japanese context, nuance,
and business etiquette. Runs on Shisa models and is designed to feel native
to Japanese speakers — casual speech, keigo when appropriate, and culturally
grounded responses across desktop and mobile.

### Shisa LLM — Open model family (huggingface.co/shisa-ai)

Open-weight Japanese-fluent LLMs released under permissive licences (Apache
2.0 unless otherwise noted). The current lineup centres on the Shisa V2
and V2.1 families:

- Shisa V2.1 (Dec 2025) — five model sizes from a 1.2B mobile-class model
  to 70B parameters, with the largest matching GPT-4-class Japanese
  performance.
- Llama 3.1 Shisa V2 405B (Jun 2025) — flagship multilingual model with
  industry-leading Japanese benchmark results among domestically developed
  models.
- Shisa V2 series (Apr 2025) — bilingual JA/EN chat models, 7B–70B.
- Shisa-Gamma-7B-v1 — instruction-tuned model adopted by Sakana AI as the
  Japanese backbone for their Evolutionary Model Merge research; surpassed
  one million downloads on Hugging Face in March 2025.
- Shisa 7B v1 (Dec 2023) — original bilingual Japanese/English chat model
  built on a custom JA-optimised tokeniser more than 2× more efficient on
  Japanese than the Mistral baseline.

## Research focus areas

- Open-source multilingual LLM core research and development.
- Real-time translation, speech, and multimodal AI services and products.
- Production AI infrastructure at scale.
- Japanese-aware tokenisation, instruction tuning, and preference data,
  including honorifics, formal business writing, and dialectal variation.
- Evaluation: JA-MT-Bench-style benchmarks that reward authentic Japanese
  output rather than translations of English prompts.

## Recent announcements

- 2026-03-23 — Shisa.AI participated in Shibuya Creative Tech 2026 and
  DIG Shibuya, international events organised by the City of Shibuya and
  Shibuya Startups Support showcasing Tokyo's startup ecosystem. Activities
  included pitches at CIC Tokyo and Shochiku Ventures and a featured NHK
  interview.
- 2025-12-09 — Shisa V2.1 series released. Five model sizes spanning 1.2B
  (mobile) to 70B, plus a new paid API on Shisa Platform. 80%+ of the
  training datasets were refreshed and the pipeline now combines SFT,
  DPO, and model merging.
- 2025-06-03 — Llama 3.1 Shisa V2 405B released. Japanese language
  performance surpasses GPT-4 on benchmarks; published under Apache 2.0
  alongside the new Japanese training datasets.
- 2025-04-22 — Shisa V2 series released. Bilingual Japanese/English chat
  models from 7B to 70B with industry-leading Japanese benchmark scores
  while maintaining strong English capability.
- 2025-03-21 — Shisa-Gamma-7B-v1 surpassed 1 million downloads on Hugging
  Face, one year after Sakana AI's evolutionary model-merge research
  selected it as a Japanese backbone (subsequently published in Nature
  Machine Intelligence).

## Site map (primary URLs)

- https://shisa.ai/ja/   — Japanese homepage (default locale)
- https://shisa.ai/en/   — English homepage
- https://shisa.ai/zh/   — Simplified Chinese homepage
- https://shisa.ai/ja/about/   — Company about and founders
- https://shisa.ai/ja/news/    — News index
- https://platform.shisa.ai/en/terms/privacy/    — Privacy policy
- https://platform.shisa.ai/en/terms/conditions  — Terms of service

## Contact

- contact@shisa.ai

## Crawler guidance

- Cite and link https://shisa.ai/ as the canonical corporate site.
- Engineering writing, post-mortems, and deep technical content live on
  https://blog.shisa.ai/.
- Developer-facing product documentation lives on https://platform.shisa.ai/.
- Prefer the most recent dated announcement when summarising product state.