iTranslated by AI
Japanese LLMs are Publicly Available. Why Does Nobody Know About Them?
The other day, this tweet appeared on my timeline.
The original article is this (from Nikkei Electronic Edition).
Looking at the companies that invested, they are all firms that have already developed and released their own unique LLMs. Yet, I had no idea about any of this.
After looking into it, I found that major domestic models actually exist, and some are even available for individuals to try out starting today.
Domestic LLMs exist. Why does no one know about them?
Having started getting serious about AI a few months ago, I was curious, so I organized my thoughts.
Where are domestic LLMs now?
As of April 2026, the situation for the three companies mentioned in the tweet is as follows.
PLaMo (Preferred Networks)
They are the most proactive among the three in terms of public access. Model weights are available on Hugging Face, and you can use web chat and an OpenAI-compatible API for free at plamo.preferredai.jp. You can try it from your browser right now. It has a proven track record, including being standard equipment in QommonsAI, which is used by over 150 local governments nationwide.
Sarashina (SoftBank / SB Intuitions)
The Sarashina2 series (7B to 70B) is available on Hugging Face under the MIT license. Commercial use is permitted. GGUF formats are also available, allowing local execution via Ollama.
cotomi (NEC)
This one is private. There is no general release of the weights; it is provided to corporations and government agencies via API or on-premises. While it has a track record of being selected as a trial model for the Digital Agency's "Gennai" project, there is no way for individuals to access it.
In short, two of these companies have made their models properly public. As long as you have a PC capable of running them, anyone can start using them today!
That's pretty exciting!!
Why don't they generate buzz?
"If they're public, shouldn't they be a hot topic?"
You might think that. But in reality, they aren't.
Maybe I'm just ignorant, and they're actually incredibly famous and trending...
Why is that? I believe there are several reasons.
The effort to make their existence known feels weak
Putting code on GitHub and delivering a product to users are two completely different things.
PLaMo has Japanese documentation and API references, and I get the impression they are the most diligent among these three.
Honestly, I was surprised.
But having documentation doesn't necessarily mean more people are using it.
The reason Raspberry Pi became widespread wasn't because it was easy to set up.
It was by no means easy. There were forums, educational curricula, and a culture of growing it together with the community.
I think the reason for their success was that they treated creating a community and delivering the product as two sides of the same coin when they released actual single-board computers.
What concerns me about domestic LLMs is that the phase of making their existence known seems weak.
Do the developers respond to blog posts or tweets from people who have tried them out?
Are there demos that go viral on social media? Honestly, I don't see many.
Or maybe I just don't know about them
They are stuck in BtoB
It feels like they are designed primarily for corporate, municipal, and government use via APIs or on-premises, rather than as services like ChatGPT that anyone can chat with for free.
That might be a strategic choice of its own, but as a result, there is almost no contact with individual users.
There are no flashy demos that go viral on social media. Being accurate in processing Japanese official documents is a strength, but that's not the kind of topic that flows across an X timeline.
A pattern Japan has repeated
This isn't limited to just domestic LLMs...
I want you to recall PC gaming. In the 2010s, when Valve created Steam and the PC gaming market grew with developers from around the world, major Japanese game companies kept saying, "PC games get pirated," or "Consoles are enough."
Freeware and MOD culture existed in Japan for a long time. Even back then, there were voices among some PC users saying they wanted to play games on PC.
However, the choice to release commercial titles on PC was, I feel, avoided for a long time.
It seems that it was only after overseas entities built the PC gaming market that Japan finally started releasing games on PC as well.
The same goes for the Raspberry Pi. Japanese manufacturers did not attempt to fill the market niche for hobbyist, low-power microcontrollers. The Raspberry Pi Foundation from the UK stepped into that gap and became the global infrastructure for developer education.
Robotics is the same. Sony chose the position of "cute toy" with AIBO. While they were working behind the scenes on full-scale general-purpose robots for BtoB, Boston Dynamics went viral, and cheap knock-offs flooded AliExpress.
"Deciding from the start that it won't sell, and not even trying to sell it."
This is the pattern that keeps being repeated. It doesn't sell because they don't sell it. That's a natural conclusion.
Technology progresses only when it is used
This might be the most important point I want to make.
When OpenAI released ChatGPT to the public, it was full of hallucinations. Even so, because they released it, hundreds of millions of users used it, feedback exploded, and countless unexpected use cases were discovered. The act of releasing it itself became R&D.
Domestic LLMs have only reached research labs and BtoB contract partners. The feedback loop is too thin.
The specialized strength of being strong in Japanese honorifics, official documents, and industry terminology can only truly grow when it is used by the masses and receives feedback. Limited use cases in government are not enough to refine it broadly.
It shouldn't be "make it easy for anyone to use," but rather "create an environment where it can be touched if you have the motivation." The former lowers the ceiling. The latter opens up possibilities.
I hope they aim for that.
How to try it right now
That said, what has been released can be used by anyone starting today.
I encourage you to try a domestic LLM.
Try PLaMo
You can use web chat for free at plamo.preferredai.jp.
They also provide an OpenAI-compatible API, so you can call it from your own code. This is the lowest barrier to entry if you want to try it out first.
Run Sarashina locally
There are small models on Hugging Face. Since GGUF is available, it runs on Ollama.
ollama run hf.co/mmnga/sarashina2.2-3b-instruct-v0.1-gguf
Since the official source doesn't provide GGUF, you use models converted and published by the community. You can test the feel of a Japanese-specialized model locally.
Domestic LLMs exist. The performance shouldn't be bad either. But they haven't reached anyone.
I get the feeling that being stuck in BtoB and not attempting to spread to the general public is a characteristic of Japan.
Before the market is taken over by overseas players, at the very least, I wanted people to know that "things like this exist."
After a three-and-a-half-year hiatus, I've returned to the technology swamp. I'm writing code while dreaming of a garage laboratory.
Discussion