Ubitus 405B

It has brought to our attention that Ubitus has continued to make ongoing misleading claims and failing to give proper credit to Shisa.AI’s work on our Shisa V2 405B model.

We find this behavior extremely disappointing, somewhat insulting, and quite frankly, more than a little baffling. We are a small startup with limited time and resources, so we have to pick our battles, and while we obviously can’t control what anyone else chooses to say or do, we can at the very least, seek to clarify the facts of the matter.

Having conducted all aspects of model training including but not limited to: training recipe creation and dataset curation, fully executing and managing the training run itself, developing and running evaluations, writing of documentation and results, and development of all custom code used (all done independently, over the course of over 4 months), we are confident that Shisa.AI is the sole author of the Shisa V2 405B model. And, while we are appreciative of the compute resources provided by Ubitus (to our understanding during development, funded in part by a METI GENIAC grant), we must also make it clear that Shisa.AI has never signed or otherwise agreed to an IP licensing or sharing agreement with Ubitus and each organization retains the respective rights to their work.

To put it succinctly, we at Shisa.AI 100% stand behind our work and claims of authorship, and encourage everyone to give a careful read and comparison of what we’ve published about our model, including:

• Our official model announcement Shisa V2 405B: Japan’s Highest Performing LLM
• Our full Overview Report (Available in English and Japanese)
• Our released model and model card at https://huggingface.co/shisa-ai/shisa-v2-llama3.1-405b

Additional Shisa.AI supporting documentation and resources:

• Over 200 ablation models, all public and dated: https://huggingface.co/shisa-ai/models
• Fully public timestamped Weights and Biases logs for all our Shisa V2 models and additional public logs for our OpenRLHF SFT and DPO runs
• Also our Overview Report (with full revision history) showing not only our authorship in both Japanese and English, that was clearly used by Ubitus and entered into the public record without due credit. (Why Ubitus would seek to highlight this is truly mystifying to us).
• After ongoing evasiveness from Ubitus in our private communications back in May, we created a private archive of records and correspondence that is GPG signed and OpenTimestamp signed that show a clear timeline and incontrovertible proof of our sole authorship of all Shisa V2 models, as well as conversations detailing the conditions under which we agreed to undertake the training of the Shisa V2 405B model.
• When was made clear that Ubitus would not be properly crediting us for our work, we proactively very publicly released our Shisa V2 405B model on June 3, prior to the June 9 METI presentation, to avoid any public confusion of authorship. We also have privately advised Ubitus multiple times to avoid making any claims that might be trivially disproved as matters of public record. At that point we would have considered the matter settled, but it seems Ubitus has continued to make additional misrepresentations.

For reference, here are Ubitus public claims that we’ve been made aware of:

• (WEB ARCHIVE 2025-06-23 Snapshot) Ubitus Unveils METI-Supported 405B Tourism LLM: Japan’s Highest-Performance Multilingual Model
  • A careful read of this article shows that Ubitus does not claim authorship or even specify the 405B model released, but omits any mention of Shisa.AI and uses the benchmarks that we conducted on the Shisa V2 405B model (raw results at shisa-ai/shaberi and shisa-ai/ja-mt-bench-harness but attributes it as “Llama 3.1 405B” (?!?).
  • This article also makes several factually incorrect claims that do not accurately represent our model (or in some cases, any Llama 3.1 model!), creating unnecessary confusion. We leave it to readers to draw their own conclusions regarding how these fundamental errors might reflect upon the credibility of any other claims presented or implied.
  • There is also a Japanese version: (WEB ARCHIVE 2025-06-22 Snapshot) ユビタス、経済産業省支援の405B観光向けLLMを発表：日本最高性能の多言語モデル
• GENIAC成果報告会 - Ubitus gave a presentation that referenced results from Shisa.AI’s model (and in fact included included tables and charts that were directly copy and pasted from the report we authored, which was shared with Ubitus under the understanding that we would be publicly credited for our work). Note, beyond the report itself and the raw results from which they were generated from, we of course have the matplotlib generation code for the graph as well as the raw evaluation results in our working repositories.