Rakuten launched 'Japan's largest AI model' with government backing. It was a fine-tuned DeepSeek V3 with the MIT license deleted. The community caught it in four hours.
Get new posts delivered to your inbox. No spam, unsubscribe anytime.
On March 17, 2026, Rakuten Group held a press conference to announce what it called "Japan's largest high-performance AI model." Rakuten AI 3.0 — 700 billion parameters, optimized for Japanese, developed under the Japanese government's GENIAC program with public subsidies. The press release was triumphant. The media coverage was glowing. The model was supposed to prove that Japan could compete with OpenAI, Google, and the Chinese AI labs on its own terms.
It took the open-source community about four hours to blow the whole thing apart.
When developers downloaded Rakuten AI 3.0 from Hugging Face and opened the config.json file, they found two fields that told the entire story: "architectures": ["DeepseekV3ForCausalLM"] and "model_type": "deepseek_v3". The technical specifications — hidden size of 7,168, intermediate size of 18,432, 61 hidden layers, 256 routed experts, vocabulary size of 129,280 — matched DeepSeek V3's architecture exactly. Not approximately. Exactly.
Rakuten AI 3.0 was a fine-tuned DeepSeek V3. And DeepSeek's MIT license file was missing from the repository.
Let me be precise about the sequence of events, because the details matter.
Step 1: Take an open-source model. DeepSeek V3 is a 671 billion parameter Mixture of Experts model with 37 billion activated parameters per token. It was released under the MIT license, which is one of the most permissive licenses in open source. You can use it commercially. You can modify it. You can build proprietary products on top of it. The only requirement that actually matters? You must retain the original copyright notice and license statement.
That's it. One requirement. Keep the license file.
Step 2: Fine-tune it for Japanese. This is normal and completely legitimate. Companies fine-tune open-source models for specific languages, domains, and use cases every day. Rakuten used computing resources provided through Japan's GENIAC project — a government-backed initiative where METI and NEDO provide access to GPU clusters for domestic AI development — to train the model on Japanese data.
Step 3: Delete the license file. This is where it goes off the rails. When Rakuten published RakutenAI-3.0 on Hugging Face, the repository did not contain DeepSeek's MIT license file. The one file the MIT license requires you to keep was the one file they removed.
Step 4: Don't mention DeepSeek. Anywhere. Rakuten's official press release made zero mention of DeepSeek. Not in the title. Not in the body. Not in the technical description. The only reference to the model's origins was a vague statement about "incorporating the essence of the open-source community." The model was promoted as a Japanese achievement, developed as part of a national project, positioned as proof that Japan could build world-class AI.
Step 5: Slap on Apache 2.0. Rakuten released the model under the Apache 2.0 license — a different, more corporate-friendly open-source license — without acknowledging that the underlying model was already licensed under MIT by its actual creators.
The discovery wasn't some elaborate forensic analysis. It was a config file.
Anyone who downloaded the model and looked at the architecture declaration in config.json could see it was DeepSeek V3. The model type, the architecture class name, the exact parameter configuration — it was all right there in plain text. You didn't need to compare weight matrices or run benchmark signature analysis. You just needed to read the metadata.
Within hours of the March 17 launch, developers on X (Twitter), Hugging Face discussion threads, and Japanese tech forums were posting screenshots of the config file. The reaction was immediate and brutal.
One widely shared comment, translated from Japanese: "Using DeepSeek is one thing, but trying to hide it is really, really pathetic."
The technical evidence was damning:
| Specification | Rakuten AI 3.0 | DeepSeek V3 |
|---|---|---|
| Total parameters | ~671B | 671B |
| Activated parameters/token | 37B | 37B |
| Architecture class | DeepseekV3ForCausalLM | DeepseekV3ForCausalLM |
| Hidden size | 7,168 | 7,168 |
| Hidden layers | 61 | 61 |
| Routed experts | 256 | 256 |
| Vocabulary size | 129,280 | 129,280 |
Every single architectural parameter was identical. This wasn't "inspired by" or "based on the same research." This was the same model with Japanese fine-tuning on top.
Once caught, Rakuten's response made things worse.
Rather than updating the repository with DeepSeek's original MIT LICENSE file — the standard, expected remediation — Rakuten added the license text under a file named "NOTICE." Not LICENSE. Not LICENSE-MIT. NOTICE.
This is a meaningful distinction. In open-source convention, a LICENSE file is the primary license declaration. A NOTICE file is supplementary information — often used in Apache 2.0 projects for attribution of third-party components. By putting DeepSeek's MIT license in a NOTICE file rather than a LICENSE file, Rakuten technically complied with the letter of the MIT license (which says you must include the copyright notice and permission notice "in all copies or substantial portions of the Software") while still keeping Apache 2.0 as the model's primary license.
The community called it out as "lacking sincerity." Legally, maybe it passed. Ethically? Not even close.
Rakuten never issued a formal apology. They never updated their press release to acknowledge DeepSeek. They never publicly explained why the license was missing from the initial release.
This scandal isn't just about license compliance. It's about public money.
Rakuten AI 3.0 was developed as part of GENIAC (Generative AI Accelerator Challenge), a program run by Japan's Ministry of Economy, Trade and Industry (METI) and the New Energy and Industrial Technology Development Organization (NEDO). GENIAC provides selected companies with access to high-performance GPU clusters and subsidies for computing costs. The explicit goal: strengthen Japan's domestic generative AI capabilities.
The Japanese government has committed approximately ¥10 trillion ($65 billion) through 2030 for AI and semiconductor infrastructure. This investment is driven by a genuine strategic concern. Rakuten's CEO Mickey Mikitani has publicly warned that AI utilization in Japan is only around 20% of the United States, and that Japan risks being left behind in the global AI race.
So the Japanese government gave Rakuten subsidized GPU access to build a domestic AI model. Rakuten used those resources to fine-tune a Chinese model. Then they presented the result as Japanese innovation without crediting the Chinese lab that built the foundation.
The public reaction in Japan was predictable and deserved. As multiple reports noted, Japanese citizens felt it was "emotionally unacceptable" that government-subsidized computing resources were used to rebrand a Chinese model as a domestic achievement.
This touches a nerve in Japan specifically because of the DeepSeek sovereignty concerns that have been simmering since early 2025. The Japanese government issued advisories about DeepSeek's data storage practices in February 2025, noting that data processed by DeepSeek is stored on Chinese servers and subject to Chinese legal jurisdiction. During Diet deliberations, lawmakers raised the issue of DeepSeek describing the Senkaku Islands as "China's inherent territory." Prime Minister Ishiba specifically stated the government would accelerate AI risk legislation in response.
Against that backdrop, one of Japan's most prominent companies quietly basing its flagship AI model on DeepSeek — and hiding it — reads as something between irony and negligence.
Here's something most coverage missed. Rakuten AI 3.0 isn't even the first Rakuten AI model built on someone else's work.
RakutenAI-7B, released in March 2024, was built by "continually training Mistral-7B-v0.1." RakutenAI-2.0, an 8x7B MoE model released in February 2025, was based on the RakutenAI-7B foundation (itself derived from Mistral). Both were released under Apache 2.0.
The difference? With RakutenAI-7B, Rakuten published an academic paper on arXiv that clearly stated the model was an extension of Mistral-7B. The attribution was explicit. The methodology was transparent. Nobody had a problem with it.
With RakutenAI-3.0, they took the opposite approach. No paper. No attribution. No mention of DeepSeek. Same company, radically different behavior with a higher-profile launch.
What changed between RakutenAI-7B and RakutenAI-3.0? The stakes got higher. The government money got bigger. The narrative of "Japanese AI independence" got louder. And apparently, the pressure to deliver something that looked like an original contribution got too intense to bother with honest attribution.
Rakuten got caught. But they're not alone.
According to Nikkei, six of the ten most prominent AI models from Japanese companies are secondary developments based on DeepSeek or Alibaba's Qwen. The Japanese AI startup Abeja, for example, uses Qwen as the base for its own models.
This isn't unique to Japan either. Chinese open-source models — particularly DeepSeek and Qwen — now account for approximately 15% of global AI model market share as of late 2025, up from about 1% a year earlier. That growth is driven almost entirely by companies fine-tuning these models for specific languages, industries, and markets. Stanford HAI has documented this pattern as a defining feature of China's open-weight AI ecosystem.
The practice of fine-tuning open-source models and releasing them under new branding is normal. The practice of stripping the original license and hiding the provenance is not.
But here's the uncomfortable question: how many other companies are doing exactly what Rakuten did and just haven't been caught?
Open-source AI model provenance is shockingly hard to enforce. The tools for detecting whether a model is derived from another are still primitive — you can compare architecture configs, run benchmark signature analysis, or compare weight distributions, but there's no automated system that flags license violations at scale. Hugging Face model cards have a base_model metadata field, but it's self-reported. If someone doesn't declare the base model, nobody checks.
The term "copyright laundering" has emerged to describe an adjacent problem: using AI to rewrite open-source code in a way that strips away provenance, attribution, and license obligations. As Sean O'Brien of the Yale Privacy Lab put it, LLMs function as "copyright removal devices — copy open-source data into it, and you get copyright-free data on the other side."
With AI models, it's even simpler. You don't need AI to launder the provenance. You just delete one file.
Here's the paradox that the open-source community needs to reckon with.
The MIT license is designed to be maximally permissive. You can do almost anything with MIT-licensed code or models — commercialize them, modify them, sell them, embed them in proprietary products. The only obligation is retaining the copyright notice.
DeepSeek chose MIT specifically to encourage adoption. And it worked. DeepSeek V3 and its variants are being used as the foundation for AI systems worldwide. The permissiveness is a feature, not a bug.
But the same permissiveness that makes MIT useful makes it trivially easy to violate. The "punishment" for ignoring an MIT license is... what, exactly? There's no automatic enforcement. There's no license compliance bot. There's no DMCA-equivalent for open-source license violations. DeepSeek could theoretically sue Rakuten for copyright infringement, but a Chinese AI lab suing a Japanese corporation over a license file? The geopolitical complexity alone makes that unlikely.
The open-source ecosystem works on trust. When someone releases code under MIT, they trust that users will honor the one condition the license asks for. When a company like Rakuten — publicly traded, government-backed, positioned as a tech leader — violates that trust, it sends a signal: the social contract of open source is only as strong as the people who choose to honor it.
Compare this with more restrictive licenses:
| License | Obligations | Enforcement Level |
|---|---|---|
| MIT | Keep copyright notice | Minimal — trust-based |
| Apache 2.0 | Keep notice, state changes, include NOTICE file | Minimal — slightly more documentation |
| GPL/AGPL | Share derivative source code under GPL | Strong — copyleft creates legal pressure |
| Llama Community License | Attribution + MAU restrictions | Corporate enforcement by Meta |
| DeepSeek Model License (pre-MIT) | Attribution + use restrictions | Moderate — custom terms |
The models with the most adoption — MIT and Apache 2.0 — are also the easiest to violate without consequences. This is the fundamental tension in open-source AI: maximizing adoption requires minimizing restrictions, but minimizing restrictions maximizes the opportunity for bad actors to exploit the system.
Lost in the licensing scandal was another problem: Rakuten's benchmark claims.
Rakuten promoted RakutenAI-3.0 as outperforming GPT-4o on Japanese benchmarks. The numbers: 8.88 on Japanese MT-Bench versus GPT-4o's 8.67. On JamC-QA, the model scored 76.9 versus GPT-4o's 74.7. On MATH-100 (Japanese), 86.9 versus 75.8. On M-IFEval (Japanese), 72.1 versus 67.3.
These are real numbers. But they're cherry-picked.
Rakuten evaluated the model on benchmarks that specifically test Japanese cultural knowledge, history, and instruction-following — exactly the tasks where a Japanese-fine-tuned version of a powerful base model would naturally excel. Nobody claimed it beats GPT-4o across the board. But the press release was structured to leave that impression.
Here's what makes this particularly dishonest: these benchmark results aren't really a testament to Rakuten's engineering. They're a testament to DeepSeek V3's quality as a foundation model. DeepSeek built a 671B-parameter MoE model that's so good it can be fine-tuned to beat GPT-4o on domain-specific tasks. Rakuten did the fine-tuning. But the hard work — the architecture design, the pre-training, the optimization — was DeepSeek's.
Taking credit for those benchmarks without crediting the foundation is like buying a Ferrari, painting it a different color, and bragging about how fast your "custom car" is.
The Rakuten scandal is a stress test for open-source AI, and open source is failing it.
The provenance problem is unsolved. There's no reliable, automated way to verify whether a published model is a derivative work. Hugging Face model cards rely on self-reporting. The base_model metadata field is optional. Weight comparison tools exist in research settings but aren't deployed at platform scale. Until model provenance is as traceable as Git commit history, this will keep happening.
Corporate incentives point toward obfuscation. When government subsidies, investor expectations, and national pride are on the line, the incentive to present a fine-tuned model as original research is enormous. Rakuten isn't some scrappy startup trying to raise a seed round. It's a $12 billion public company. If Rakuten's compliance team let this happen, imagine what's happening at smaller companies with less scrutiny.
License enforcement is basically non-existent. The MIT license is the honor system of software licensing. It works when the community is small and people know each other. It works less well when $65 billion in government subsidies are on the table. The open-source community needs to decide whether social pressure alone is sufficient enforcement, or whether it's time for something stronger.
The "sovereign AI" narrative creates perverse incentives. Japan, France, the UAE, Saudi Arabia, India — governments worldwide are investing in domestic AI capabilities. The political pressure to show results is immense. When the fastest path to a "national AI model" is fine-tuning DeepSeek and deleting the license file, some companies will take that path. The Rakuten scandal won't be the last.
Rakuten's behavior was indefensible. Not because they used DeepSeek V3 — that's exactly what open-source AI is for. But because they actively tried to hide it.
The MIT license asks for one thing. One thing. Keep the copyright notice. Rakuten couldn't even do that. And when they got caught, instead of a real mea culpa, they buried the attribution in a NOTICE file and hoped the news cycle would move on.
I think this scandal reveals three things:
First, "sovereign AI" is mostly a fiction. The idea that Japan (or France, or India, or any country except the US and China) is going to train globally competitive foundation models from scratch is unrealistic. The compute costs alone — DeepSeek V3 cost an estimated $5.5 million to train, and that's considered cheap — put from-scratch training out of reach for most national programs. Fine-tuning open-source models IS the realistic path to useful domestic AI. There's no shame in it. The shame is in pretending otherwise.
Second, the open-source AI ecosystem needs better provenance tooling. We need the equivalent of SBOM (Software Bill of Materials) for AI models. Every published model should have a machine-readable declaration of its base model, training data sources, and license chain. Hugging Face is best positioned to build this, and they should make it mandatory, not optional.
Third, trust is the only currency that matters in open source. DeepSeek released their model under MIT because they wanted the world to build on their work. Rakuten abused that trust. If this becomes a pattern — and the Nikkei data suggests it already is — it will push model developers toward more restrictive licenses. That's bad for everyone. The solution isn't more restrictive licensing. It's enforcing the licenses we already have.
The saddest part? If Rakuten had simply written "Rakuten AI 3.0 is a Japanese-optimized fine-tune of DeepSeek V3, developed with government support through the GENIAC project" — that's a perfectly respectable achievement. Fine-tuning a 671B model to beat GPT-4o on Japanese benchmarks is real engineering work. But by hiding the provenance, they turned a legitimate accomplishment into a scandal.
Open source runs on trust. Rakuten spent theirs.