79 state-backed AI models, 1/90th the downloads of one Qwen: a report card for Japan's GENIAC program

Japan's government helped pay the computing bill for a wave of homegrown AI models. So how often do people actually download them?

GENIAC is a program run by Japan's Ministry of Economy, Trade and Industry (METI) and its research-funding agency NEDO. Since 2024 it has selected 70 AI development projects. We traced 79 models on Hugging Face, the main site for sharing AI models, back to GENIAC using the program's official results pages and the models' own documentation, and counted their downloads.

Over the last 30 days, all 79 together were downloaded about 78,000 times. Alibaba's Qwen3-8B was downloaded about 7 million times on its own. Combined, the 79 models get roughly 1/90th of the downloads of a single popular overseas model.

That number alone doesn't mean nobody is using them, though. PLaMo 2.0 Prime, a model Preferred Networks (PFN) built in GENIAC's second round, was already running inside an AI assistant used by more than 150 Japanese local governments when it launched; its successor now runs in about 800. Kotoba Technologies, which built speech models through GENIAC, says more than 180,000 people use its live-interpretation app. Much of Japan's homegrown AI is used inside services, not downloaded.

WHAT IS

GENIACGenerative AI Accelerator Challenge

A Japanese government program that subsidizes computing costs to build domestic AI foundation models

1. 2024.02Launch, 10 projectsMETI starts the program and picks universities and companies to build large Japanese-language models.
2. 2024.10Round 2: 20 projectsA ¥24.5 billion cap. More models aimed at specific fields such as drug discovery, anime and climate.
3. 2025.07Round 3: with users24 projects. Trials with user companies are added, and Rakuten builds a giant model.
4. NOWRound 4: niche models16 projects picked in June 2026. Renamed, with extra points for trials with user companies.

GENIAC is a program METI and NEDO started in February 2024 to strengthen Japan's ability to build generative AI. Its core is a subsidy for the cost of the GPUs (the chips used for AI computing) needed to train models. The launch announcement put "support for providing computing resources for foundation model development" at the center.

So far it has selected 10 projects in round 1, 20 in round 2, 24 in round 3 and 16 in round 4, for 70 in total, out of more than 150 applications. Across five calls for proposals, the budget caps add up to ¥53.9 billion (roughly $360 million at ¥150 to the dollar).

The list of recipients includes companies such as Rakuten, PFN, ABEJA, Stockmark, Sakana AI, KARAKURI, Ricoh and Nomura Research Institute, as well as the University of Tokyo and the National Institute of Informatics. Summarizing a Japanese meeting, reading a contract, interpreting a conversation in real time: the goal is to make sure the models behind everyday features like these can be built in Japan.

GENIAC is now shifting from building to being used. From fiscal 2026, proposals that include trials with user companies get extra points, and the program has a new name, the "Domain-Specific Model Development Program." Japan's Digital Agency, meanwhile, has government employees comparing domestic AI models with the source hidden, in tests running through November. A summary of the results is due this winter.

In March 2026, a different question went viral on X. Rakuten AI 3.0, a model built with GENIAC support, turned out to be built on a publicly released model from China's DeepSeek. On November 5, Rakuten added that history to the model's own documentation. What should "homegrown AI" mean when public money pays for it? The question is back.

Are the models being used? As a newsroom that works with AI and data every day, we wanted an answer in numbers. Using government documents, company announcements, other countries' programs and Hugging Face data retrieved on the night of November 7, 2026, we graded GENIAC.

Building a large model of the kind behind ChatGPT means running huge numbers of GPUs for months. In 2023, few Japanese companies or universities could line up that much computing power on their own.

So the government decided to subsidize the computing bill. METI's R&D plan explains why:

Whether or not Japan has the capability to develop foundation models may determine how available generative AI is in Japan and how broad the innovation it creates can be. ── METI, R&D plan for the Post-5G program (revised November 5, 2026)

A country that can only use foreign AI leaves prices and terms of use in other companies' hands. GENIAC started from the idea that Japan should keep the ability to build AI that fits its language and the way its businesses work.

The subsidy covers GPU costs and, from round 2, outsourced work to prepare training data. It pays two-thirds of the cost for small companies and universities and half for large companies, which cover the rest themselves. The state pays part of the computing bill rather than funding development outright.

From general-purpose models to specialists

What GENIAC funds has changed with each round. Round 1 was about simply building a general-purpose Japanese model first. A 172-billion-parameter model from the National Institute of Informatics (parameters are a rough measure of a model's size), a 100-billion-parameter model from PFN, and "Tanuki," built by the University of Tokyo's Matsuo-Iwasawa Lab with about 250 volunteers, all came out of this period.

From round 2, models for specific fields multiplied: drug discovery, anime, climate, pharmaceuticals, tourism, document reading. Round 3 added trials with user companies, and round 4 targets even narrower areas, such as security cameras, finance, car maintenance and materials. At the round-3 results briefing, NEDO's department head described round 1 as a stage of "let's just try building a foundation model in Japan," followed by industry-specific models and then trials with users.

"Homegrown" includes models built on foreign ones

There's another thing to settle: what counts as "homegrown." GENIAC supports not only models trained from scratch but also extra training of existing models, including foreign ones, in Japanese and other skills. NEDO's own definition of a domestic foundation model includes "models created by additional training of an existing foundation model (not limited to domestic ones)."

That line drew a lot of attention in March 2026. Rakuten AI 3.0, which Rakuten released on March 17, is Japan's largest model at about 700 billion parameters, built with GENIAC round-3 support. On launch day, a developer who goes by kis asked on X whether, given its size, it was DeepSeek-V3 with extra training. Posts making the same point were viewed tens of thousands of times.

Rakuten first declined to name the base model. In a press response dated March 27, it confirmed it had used DeepSeek's open model. Then on November 5, it added release notes to the model card explaining that DeepSeek-V3-0324 was further trained on English and Japanese text, and that the result beat its base on questions about Japanese culture but fell behind it in math.

Extra training to boost a model's Japanese is a common technique worldwide. The real debate was less about technology than what people expect from "homegrown AI" built with public support.

How much of GENIAC's budget is public? The often-quoted figure of ¥33.9 billion is what Nikkei xTECH reported in March 2025 as the total of the calls so far. Checked against METI's documents, it equals the budget caps of the first three calls added together (¥8.4 billion, ¥1 billion and ¥24.5 billion).

Add the two calls since then, and the five budget caps total ¥53.9 billion.

These are caps. What was actually paid has not been disclosed, either for GENIAC as a whole or per company. The money comes from the "Post-5G" fund, which also pays for chip development and received ¥2.78 trillion between fiscal 2019 and 2025. Most of it goes to semiconductors, including the chipmaker Rapidus; GENIAC is one theme within it. In the budget request for fiscal 2027, GENIAC is listed without an amount.

What got built: from scratch, or on top of foreign models

GENIAC's official site lists results for rounds 2 and 3. Round 2 produced 21 models and 9 datasets; round 3, 17 models and 10 datasets.

The range is wide.

Source: DATA TIMES, based on METI GENIAC's round-2 and round-3 results pages, results briefing reports and company announcements

Of the 79 Hugging Face models we counted, 31 had their core (the part that handles text or video) trained from scratch by the developer, and 48 were built on a public model, mostly foreign ones such as Qwen, Llama or DeepSeek. About 60% of GENIAC's published models are built on someone else's open model. We sorted them using the "base model" field in each model card and the developers' own descriptions. Models that only borrow a part for reading images, such as PFN's PLaMo 2.1 VL, which uses a Google component, count as built from scratch.

PFN's PLaMo, Stockmark's 100-billion-parameter model and the University of Tokyo's Tanuki were trained from scratch. ABEJA, KARAKURI, Ricoh, Nomura Research Institute and Turing's vision-language models started from Alibaba's Qwen or OpenAI's open models and added Japanese and industry knowledge. Pick a base and ship fast, or build from scratch and own everything inside: the companies' different strategies show up here.

One round-2 recipient had its grant fully revoked by NEDO in August 2025 after false information was found about its finances at the time of its proposal.

No yardstick yet for whether the models are used

So how does the government grade GENIAC's results? The R&D plan's target is to "verify that the generative AI foundation models developed have a sufficient advantage by international comparison." Each applicant chooses its own benchmark (the test used to measure performance).

GENIAC has no performance indicators of its own. The only ones are for the whole fund, chips included: at least 80% of projects meet their targets, and at least 50% reach practical use about three years after they end. We found no indicator for downloads, the number of companies using a model or how often it is used.

Government reviewers have flagged this. The 2023 public review of the program concluded:

The long-term outcome is set as "a practical-use rate of 50%," but "practical use" can mean many things, so the addition of an indicator of economic impact should be considered. ── Administrative Program Review, 2023 public review summary (Post-5G program)

One outside measure of use is downloads. Most openly released AI models live on Hugging Face.

We pulled the numbers through Hugging Face's public API (an interface that lets programs fetch data) between 10 p.m. and midnight Japan time on November 7, 2026. We counted 79 models that met at least one of these conditions:

• Linked from GENIAC's official round-2 or round-3 results pages
• Model documentation mentions "GENIAC" or "JPNP20017," the project number of GENIAC's funding source
• Described as a GENIAC result in the developer's official announcement

Versions repackaged to run more easily on a local machine (quantized versions and the like) count separately when they have their own Hugging Face page. Merging those, the 79 come down to 72 distinct models.

The result: 77,976 downloads in 30 days. Alibaba's Qwen3-8B alone logged 6,976,036. Meta's Llama-3.1-8B-Instruct had about 4.35 million.

The median for the 79 models was 86 downloads a month. 43 of the 79 were downloaded fewer than 100 times in 30 days.

The total is dominated by one model. PFN's "PLaMo 2.1 2B VL," a small model that can also read images, built in round 3, logged 57,290. That one model accounts for 73% of all 79. It has 62,930 downloads since it was released in March, so nearly all of them came in the past month. The reason hasn't been disclosed.

Next come AIdeaLab's "AnimeGen," which turns a drawing into an anime clip (3,112), a version of Sakana AI's small model "TinySwallow" for running locally (2,222), PFN's "PLaMo 2 1B" (1,553) and "PLaMo Translate" (1,517). Rakuten AI 3.0, for all the attention it got, was downloaded 316 times.

Compared with Japanese models built outside GENIAC

Japanese models are also being built outside GENIAC. We counted the main series at the same time.

Source: DATA TIMES tally via the Hugging Face public API (November 7, 2026). LLM-jp-4 and Swallow are models with the series name in their name; Sarashina is SB Intuitions models with "sarashina" in their name.

Swallow, which adapts Llama and Qwen to Japanese, came out at about the same scale as GENIAC's 79 models. Japanese-language models as a whole sit about two orders of magnitude below a popular overseas model.

The standout is the National Institute of Informatics' LLM-jp-4, at more than 370,000. But 278,160 of those come from a single file converted to the GGUF format, which makes it easy to run on a laptop. LLM-jp-4 was funded by the education ministry, not GENIAC. Its predecessor, LLM-jp-3, though, says in its documentation that GENIAC partly supported it. The experience gained under GENIAC appears to have carried into the next generation.

1,353 derivative models, 80% from a single model

Derivative models are another clue. These are models that others built by further training an existing model and then published, which is evidence that someone used it and reworked it.

The 79 models have 1,353 derivatives in total. 1,117 of them come from a single model, the National Institute of Informatics' LLM-jp-3 13B. They were made by 694 different people or organizations, mostly in November and December 2024.

Following the derivatives' documentation, many turn out to come from a competition in the University of Tokyo Matsuo-Iwasawa Lab's 2024 course on large language models. Students further trained this model and competed on the quality of its answers. A Japanese model built with state support became a training ground for hundreds of students and working adults.

Beyond monthly downloads, the 79 models have about 2.44 million downloads in total since release and 1,075 likes.

The use downloads can't see

These numbers have real limits.

Hugging Face's download count is the number of times files were fetched. It includes repeat fetches by the same person and automated tests, so it isn't a count of users. On the other hand, models used only inside a company, services called over the internet through an API, and models built into products don't show up at all. And a model like Rakuten AI 3.0, which requires users to accept terms before downloading and needs large servers to run, will naturally see few downloads.

So here are the numbers the companies themselves have published:

• PFN's PLaMo: PLaMo 2.0 Prime, a version of a GENIAC round-2 model tuned for serving, was in use in QommonsAI, an AI assistant adopted by more than 150 local governments, when it was announced in May 2025. Its successor, "PLaMo Prime," now comes built into QommonsAI, adopted by about 800 local governments (October 2026). PFN says "PLaMo Translate" was developed with GENIAC support, and since December 2025 it has been available to staff through the Digital Agency's government AI platform "Gennai".
• Kotoba Technologies: Built speech foundation models in GENIAC rounds 1 to 3, and announced that its live-interpretation app has passed 180,000 users (June 2026).
• Stockmark: Put what it learned training its round-2 model into its own information-gathering service (July 2025).
• Nomura Research Institute: Said at the results briefing that its round-3 model let it check sales conversations for compliance "with 100% coverage," up from about 5% before (the company's own account).

Rakuten AI 3.0, by contrast, had not yet been put into Rakuten's AI services as of March, according to reports, and no deployments have been announced since.

Backing your own country's AI, often called "sovereign AI," is now a policy goal in many places. The easiest comparison for GENIAC is South Korea.

In August 2025, Korea's Ministry of Science and ICT picked five teams to build "sovereign AI foundation models" and narrows the field after each evaluation. The government makes training from scratch, without another model's learned weights, a minimum requirement. In the first evaluation in January 2026, even NAVER's team, which ranked near the top overall, was dropped for failing that test of independence.

In the second evaluation in August 2026, Upstage, SK Telecom and LG AI Research advanced to the next stage. Benchmarks were worth 40 points, experts 35, and ratings from people who actually used the models 25. 185 members of the public joined the user panel.

Korea's own models don't get many downloads either. A tally by Korean IT outlet Digital Daily in August put last-month Hugging Face downloads at more than 26,000 for Upstage's Solar Open 2, about 3,000 for SK Telecom's A.X K2 and just over 1,600 for LG's K-EXAONE 2.0. That's the same order of magnitude as GENIAC's results. The difference is that Korea scores whether the models are used and creates places for them to be used.

Other countries have their own approaches:

• India: The IndiaAI Mission has a five-year budget of about ₹103.7 billion and chose 20 foundation-model projects from 506 applications. It rents GPUs out at subsidized rates, and the developers keep the intellectual property.
• United Kingdom: The Sovereign AI Unit invests in promising AI companies with up to £500 million. Rather than building a national language model, it grows companies with computing power and equity.
• France: The government's digital agency built an AI tool for all state employees together with Mistral AI (June 2026), making the state the first big customer.
• Singapore: SEA-LION, a model for Southeast Asian languages, is built by further training foreign models such as Qwen and Gemma.
• Switzerland: Universities and the national supercomputing center built Apertus, releasing everything down to the training data.

Set side by side, three differences from GENIAC stand out. Korea narrows the field, publishes scores and lets users help decide. Korea and France create places where the state uses the models. And "homegrown" means very different things, from Korea's from-scratch requirement to Singapore's extra training of foreign models. GENIAC selects broadly and accepts many ways of building, an approach designed to widen the base.

Japan has started creating demand too. The Digital Agency picked 7 domestic models from 15 applications for its Gennai platform (March 2026). The only one officially confirmed as a GENIAC result is PFN's PLaMo 2.0 Prime. Since July, government staff have been trying models from five companies and comparing answers with the source hidden. And NEDO's prize competition, GENIAC-PRIZE, now requires entrants to use a domestic model from fiscal 2026.

Seventy projects in two and a half years. Several models of around 100 billion parameters were trained from scratch, and models for specific fields followed. The subsidy clearly moved development that computing costs had been holding back. On the ability to build, we'd give GENIAC a passing grade.

Whether the models are used, though, can't be judged right now. The government looks only at scores on tests the applicants chose and at the fund-wide practical-use rate; no figures on how much GENIAC's results are used have been published. Downloads, the one thing outsiders can count, add up to about 1/90th of a single popular overseas model. A program with ¥53.9 billion in budget caps should count its own usage and publish it. That alone would raise the level of the debate.

Counting made one thing clear: downloads barely capture how Japanese AI is used. PLaMo 2.0 Prime, built under GENIAC, went into an AI assistant used by more than 150 local governments, and its successor now runs in about 800. Kotoba, which built speech models through GENIAC, has 180,000 people using its interpretation app. Neither shows up in download counts.

Japanese companies and government offices want AI that keeps data in-house and reads Japanese documents correctly. So homegrown models tend to be built into services and internal systems rather than handed out in public. That means what the government should measure is how many services a model is built into and how many people use it at work. Companies already publish some of this; collect it, and it can be measured.

What we're watching most closely is the Digital Agency's Gennai test. Government staff pick the better answer without knowing which model wrote it. Like Korea's user ratings, it's the first time the people using Japanese AI get to grade it.

A summary is due this winter. How will staff rate PLaMo 2.0 Prime, a GENIAC result, next to tsuzumi 2 and Takane, which were built outside GENIAC? From fiscal 2027, the government plans to pay for the best-rated models. If the state becomes the first customer, GENIAC's report card will finally have something in the "used" column. We'll check back when the results are out.