Shisa.AI 2025 Year in Review

Happy New Year from Shisa.AI!

Dear Friends & Family,

As we kick off 2026, we wanted to take a moment both to look back at an incredibly busy first year and share what’s ahead for 2026. I’m sure this year will be a roller-coaster. Thanks for believing in us!

How You Can Help

Get Connected:

  1. IMPORTANT: Join our Discord ->> LINK. Our company updates will be coming from here going forward. There’s lots of great conversation here too. Be sure to let us know your Discord username to get access to the PRIVATE #supporters channel
  2. Email address for iOS Test Flight or Google Play - we are soon to beta test our native apps and would love to have you guys kick the tires first!

Refer Partners and Customers to us:

  1. Japanese companies that need translation tools to globalize better, LLMs and AI tools to modernize
  2. Global companies looking to deploy products/services in Japan
  3. Japanese System Integrators (SI)

Japanese Sales Hire Referrals

  1. Enterprise Sales (senior + junior roles) to drive revenue

Company Milestones

2025 marked Shisa.AI’s first year as a company!

  • January: Incorporated Shisa AI Inc. (US Delaware C-Corp)
  • May: Established Shisa株式会社 (Japan KK)
  • June-October: Friends & Family / Pre-Seed round
  • September: NVIDIA Inception, AWS Accelerate, and Google for Startups
  • Talks throughout the year presenting at Tokyo AI, Rakuten, University of Tokyo, and more
  • Featured in press, including our first appearance on national TV on the TV Asahi Boostars! program.

R&D Highlights

Our open-source mission remains core to who we are. We’ve published ~300 models and a dozen open datasets on our Hugging Face (and hundreds more private models and datasets).

We released SOTA open Japanese LLMs throughout the year. Currently, our models are the best-performing Japanese open models from 350M to 405B in size, and Shisa V2 405B is still the strongest LLM ever trained and released from Japan.

  • Shisa V2 (April)
  • Shisa V2 405B (June)
  • Shisa V2.1 (December)

Besides models and data, we’ve also made some major open source contributions:

  • megablocks-hip: 5X faster MoE training for AMD GPUs
  • Axolotl PRs - improved data handling, distributed-friendly Muon Clip Optimizer, Aux Loss Free Balancing
  • Multiple evaluation benchmarks: Shaberi V2.1, shisa-jp-ifeval, shisa-jp-rp-bench, shisa-jp-tl-bench

We closed out officially releasing JP-TL-Bench (a general-purpose Japanese-English translation benchmark that we believe is significantly better than anything that has been publicly released) and publishing our first paper on arXiv.

Of course, we have quite a bit of work that we’ve done internally as well. Our shisa-moe framework is currently training Qwen 3 30B-A3B MoE models >80% faster than the best publicly released code on our dev GPUs. Our “chotto” models are tiny and on dialogue/conversational JA/EN translation, are quality competitive with Gemini 2.5 and a fraction of the size and inference cost, and we’ve begun training improved JA/EN ASR, cutting speech recognition errors in half already.

Products Taking Shape

As most of you know, while we’re not a Translation company, this is our first target market for both strategic and tactical reasons.

chotto.chat (Real-Time Translation)

From the first commit in March to a live beta, Chotto is our flagship product.

  • We’ve iterated constantly, and the product keeps getting better
  • Average session time is 27 minutes - users love it
  • iOS and Android versions are around the corner

chotto.events

  • Specific version for conferences, talks, and networking events
  • We’ve been used at over a dozen events in November and December so far
  • This includes big tech events like Ramentech in Fukuoka, Techstars Tokyo, and TokyoAI
  • As well as non-tech events like a Sake tasting and Women Career events

Chat bots

  • Our corporate offering also includes a Slack and a LINE Chat bot for translation

Specialized Translation

  • We have begun attacking industry verticals both for revenue and for generating proprietary datasets
    • Automotive/manufacturing
    • eSports (fighting and puzzle games)
    • Dating
    • Food

Shisa Talk API

Launched in November at talk.shisa.ai:

  • Translation API (the same one powering Chotto; there’s a drop-in DeepL API replacement layer as well)
  • Voice/TTS
  • LLM Inferencing (we’ve seen demand for hosting our Open Source models!), OpenRouter integration coming soon

Sovereign AI

We continue to shape our GTM/strategy for deploying on-premise and custom training solutions for enterprises requiring data privacy and security, with translation as our foot-in-the-door.

Looking Ahead to 2026

We’re entering 2026 with momentum and clear priorities:

Scaling the Team

We’re hiring across key functions:

  • COO to help operationalize our growth
  • Enterprise Sales (senior + junior roles) to drive revenue
  • Marketing/Social Media to grow our presence in the Japanese AI community
  • LLM Researcher to accelerate our research publications

Research Roadmap

  • Shisa V3 (Omni): Our multimodal model, GENIAC Round 4
  • Improved speech recognition (ASR) and OCR capabilities
  • New conversational voice model improvements
  • More languages and new language pairs for our models
  • Japanese KMS/KB Agents with tool calling and reasoning

Product Development

  • Native Apps (soon!)
  • AR/AI Glasses and other hardware integrations
  • Enterprise deployments

Thank You

2025 was a sprint where we took an Open Source project and turned it into a startup - shipping models, yes, but launching them into products, making key hires, and signing our first clients.

2026 is about scaling. With your continued support, we’re growing our team, focused on executing, and establishing Shisa as the go-to for Japanese AI.

Here’s to an exciting year ahead!

Warm regards,

Leonard, Jia and the rest of the Shisa.AI team

Questions? Want to chat? Just reply to this email—we’d love to hear from you.