Shisa.ai
Three-person Tokyo lab, open Japanese LLMs beating far bigger models
Overview
Shisa.ai is a three-person Tokyo startup building open-weight Japanese-English LLMs (Shisa V2/V2.1, 1.2B to 405B parameters) plus translation, text-to-speech and speech-recognition APIs. Its voice-translation app Chotto draws over 6,000 daily users pre-launch.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 12 Jul 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
What you can produce with Shisa.ai
- Open-weight Shisa V2/V2.1 Japanese-English LLM family (1.2B–405B parameters)
- Translation API (Japanese-English)
- Text-to-speech API for Japanese ($0.08/1K characters, Flash Monolingual)
- Speech recognition (ASR) API
- Chotto real-time voice-translation app
- Open Japanese-language evaluation benchmarks (JP-TL-Bench, jp-ifeval, jp-rp-bench)
- Model weights and eval tools published on Hugging Face and GitHub
ASEAN Perspective
Shisa.ai in Southeast Asia
ASEAN-region availability and pricing notes coming soon. Drop the editorial team a note via /contact/ if you can supply local context (Singapore/Malaysia/Indonesia/Thailand/Vietnam).
Shisa.ai is the rare small lab whose benchmark claims hold up on inspection: its 405B model pulled over a million Hugging Face downloads in its first month, and the newer, much smaller V2.1-14B reportedly beats the older V2-70B on Japanese-language tasks. Founder Jia Shen has said data quality, not scale, was the single biggest lever — a refreshingly unglamorous explanation for what's usually sold as a scaling story.
Everything ships as open weights, which is unusual among Japanese-focused labs and genuinely useful if you want to self-host without vendor lock-in. The company is tiny, though — three people as of its last public interview — so support depth and long-term roadmap certainty are real questions for anyone building production dependencies on it. Its TTS API is priced per character with volume discounts on request, typical of a young API business still finding its enterprise footing.
What people say
Shisa.ai is co-founded by CEO Jia Shen and CTO Leonard Lin, with a third team member, anime translator Adam Lensenmayer, credited with sharpening the model's grasp of natural Japanese register — his prior work includes Attack on Titan and Detective Conan localization. All three are immigrants who settled in Japan and have framed the company's mission explicitly around AI sovereignty: building local-language models as infrastructure, not just a product line.
The technical record is unusually well-documented for a company this size. Leonard Lin and collaborators released the original Shisa 7B EN-JA model in December 2023; the follow-up Shisa V2 family scaled up to a 405B-parameter flagship that pulled over a million downloads on Hugging Face within a month of release, with independent commentary comparing its Japanese-task performance to GPT-4o and DeepSeek-V3. The V2.1 refresh, covered on the company's own blog, adds new small models at 1.2B, 3B and 8B for edge deployment and claims the 14B variant now exceeds the older 70B model on Japanese benchmarks — achieved, the team says, without benchmark-specific tuning, using stricter GPT-5.1-based judging for evaluation.
Shisa.ai has also built its own Japanese-specific evaluation suite rather than relying solely on general benchmarks: JP-TL-Bench for translation quality, Shisa-jp-ifeval for verifiable constraints like mora counting and honorific usage, and Shisa-jp-rp-bench for role-play dialogue quality. These are published as open tools on GitHub under the shisa-ai organization, alongside the model weights.
On the product side, Chotto — a real-time voice-translation app — is still pre-launch but already logging over 6,000 daily users with average sessions around 27 minutes, according to the company. Commercial API pricing is usage-based rather than tiered: text-to-speech runs $0.08 per 1,000 characters for the Flash Monolingual model, with volume discounts available on request. There's no independent third-party review data (G2, Capterra) for the API business yet — most of what's public comes from the company's own blog and press interviews, so claims about benchmark superiority should be read as self-reported until wider third-party evaluation catches up.
Summary of public user & expert reviews, compiled by RECATOOLS.
About this listing
This entry was compiled from publicly available data including Shisa.ai's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with Shisa.ai unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to Shisa.ai directly →
Spotted something out of date? Suggest an update →
Alternatives to Shisa.ai
More in LLMs & Chat
