Rakuten and HP Put Japanese AI Directly Inside PCs

Japanese personal computers have always had to decide where the language should live.

In the early 1980s, a business machine needed enough memory, processing power and display hardware to handle thousands of kanji. NEC’s PC-9801 carried Japanese capability in a system built specifically for the domestic market. Later, DOS/V and Windows moved more of that work into software running on internationally compatible hardware. The question looked technical—ROM, graphics controller, operating system—but its answer decided which machines, programs and companies could enter Japan’s computing world.

Rakuten and HP are reopening that old architectural argument in a new form. Their subject is no longer where to draw a character. It is where a machine should interpret a sentence, summarize a confidential document, understand a traveler’s intention or turn a shopping request into a commercial action.

Rakuten AI for Desktop offers two answers in one application. A compact Japanese-language model can work on the HP PC without an internet connection. A cloud mode can reach larger capabilities and agents connected to Rakuten Ichiba, Rakuten Travel, Rakuten Mobile, Rakuten Music, Rakuten Point Club and other parts of a group that spans more than 70 services.

The most consequential feature is therefore not a chat window or a floating icon. It is the boundary between the two modes. A hybrid AI PC is a machine with two places to think—and with economic, privacy and security consequences attached to the route between them.

One window, two kinds of intelligence

On-device mode runs Rakuten AI 7B ONNX, a specially trained, fine-tuned and quantized version of Rakuten’s seven-billion-parameter model. It handles three plainly defined jobs without a cloud connection: writing and rewriting text, translation and summarization. The companies emphasize low latency, offline continuity, reduced cloud expense and the privacy advantage of keeping the selected material on the PC.

Cloud mode is broader. It provides AI search, image and short-video creation, voice notes, coding assistance and problem solving. It also opens specialized Rakuten agents: product search and recommendations in Rakuten Ichiba, hotel search and booking assistance in Rakuten Travel, music discovery, Rakuten Mobile guidance, point-balance checks and golf-course search through Rakuten GORA.

A third layer is social rather than transactional. The announced menu includes an anime conversation partner, English teacher, cooking partner, life and career coach, fashion adviser, fitness coach, fortune teller and general pet-care adviser. The breadth reveals the strategic ambition. Rakuten is not presenting a single office utility; it is trying to occupy the conversational doorway through which a user reaches work, entertainment and commerce.

The interface is designed to make that doorway difficult to ignore. A central chat window connects to the agents. A floating icon can sit on the screen or in the system tray. A contextual action bar can propose a summary, translation or rewrite beside selected text, reducing the familiar ritual of opening a browser tab and pasting material into a separate chatbot.

Activation is not quite the same as traditional preinstallation. The icon on the taskbar or Start menu opens a website, the user downloads the application and then launches it. A mode switch checks whether the hardware supports the local model; if it does, the on-device component can be installed from settings. A customer may therefore own an eligible bundled HP PC yet still have only cloud mode if the machine fails the local compatibility check.

Where the task runs	Documented capabilities	What the boundary means
On the HP AI PC	Writing and rewriting, translation and summarization using Rakuten AI 7B ONNX; usable offline on compatible hardware.	The selected content can be processed locally, reducing network delay and the need to send that task to a remote service.
Rakuten cloud	Search, image and video creation, voice notes, coding, problem solving, community agents and connections to Rakuten services.	Broader knowledge and live services require connectivity and bring cloud processing, accounts, permissions and service policies into play.
Not yet disclosed	Minimum local specifications, model file size, quantization level, context window, comparative benchmarks and a detailed automatic-routing policy.	Buyers cannot yet independently compare quality, performance, storage cost or exactly how future versions may choose between local and cloud execution.

The small model is the point

Seven billion parameters sounds enormous until it is placed beside frontier cloud systems. Rakuten itself unveiled Rakuten AI 3.0 in December 2025 as an approximately 700-billion-parameter mixture-of-experts model, activating about 40 billion parameters for each token. That system belongs in substantial computing infrastructure. The older 7B line can be compressed into the practical limits of a personal computer.

Rakuten AI 7B began with Mistral-7B-v0.1, an open model from France’s Mistral AI. Rakuten continually trained it on curated Japanese and English data, created foundation, instruction and chat versions, and released them under the Apache 2.0 license in March 2024. For HP PCs, Rakuten says it performed further continual training, fine-tuning and quantization and packaged the result as Rakuten AI 7B ONNX.

Quantization represents numerical values in the model with fewer bits. It can shrink the model and reduce the memory and computation needed for inference, although the exact method can affect quality. ONNX provides a common representation that runtimes can optimize across hardware. Microsoft says Rakuten AI for Desktop uses Foundry Local, which can connect applications to Windows acceleration across NPUs, GPUs and CPUs rather than forcing the developer to engineer a separate path for every chip.

The Japanese tokenizer is as important as the parameter count. A model does not read a sentence as a human does; it breaks text into tokens. Rakuten extended its tokenizer so a token can represent more Japanese characters than before. More efficient tokenization can shorten the sequence the model must process, lowering training and inference cost and helping a small local model do useful work.

That history also supplies a needed restraint. Rakuten’s own December 2025 Japanese MT-Bench comparison scored the original Rakuten AI 7B at 4.35, Rakuten AI 2.0 at 6.79 and Rakuten AI 3.0 at 8.88. The tests, versions and purposes are not identical to the new HP implementation, so they are not a product review. They do show why hybrid design exists: the model small enough to travel in a PC is not automatically the model best equipped for every difficult question.

From kanji ROM to local language model

The resemblance to Japan’s first PC age is deeper than nostalgia. Eight-bit personal computers were poorly suited to business Japanese, whose character codes and display needs demanded more than ordinary hobby machines could comfortably provide. NEC announced the 16-bit PC-9801 in October 1982 with Japanese-language processing and color graphics. It became the center of a domestic hardware and software culture.

Its advantage was also a border. Japan’s language requirements helped sustain architectures that were not simply interchangeable with the IBM-compatible world. The Information Processing Society of Japan’s computer history records how PC-98 machines used kanji ROM and how DOS/V later enabled Japanese display on PC/AT-compatible hardware. Windows 95 accelerated the shift toward globally standardized machines with Japanese handled more flexibly in software.

The AI era repeats the movement at a higher level. A cloud model trained primarily for global scale can speak Japanese, but fluency is not the same as understanding the compressed politeness of business email, omitted subjects, era names, address conventions, service expectations or the difference between a literal translation and an appropriate one. A locally optimized model is an attempt to place those linguistic costs back inside the machine.

Yet the old history warns against easy nationalism. The PC-98’s Japanese specialization was useful; compatibility eventually mattered more. Rakuten AI 7B itself descends from a French open model, runs through an international model format and uses Microsoft’s local runtime on American-designed operating systems and multinational silicon. “Domestic AI” here describes adaptation, data, engineering and service context—not an isolated national stack.

Rakuten’s long route from marketplace to agent

Rakuten began in 1997 with a proposition that sounded doubtful at the time: small merchants could build businesses in an online shopping mall. The first Rakuten Ichiba had 13 merchants. Rakuten Travel followed in 2001 and the points program in 2002. By 2006, the company was describing an “ecosystem” in which one membership and loyalty currency encouraged movement among different services.

That history matters because an agent becomes more useful when it can do more than answer. A general chatbot can recommend a hotel. An ecosystem agent can know the inventory format, understand the loyalty program, navigate the booking flow and keep the user inside one commercial identity. Rakuten’s group now spans e-commerce, travel, cards, banking, securities, digital content and mobile communications, with more than 70 businesses and about 2.1 billion members globally by the company’s count.

The model program arrived in stages. Rakuten released the 7B family in March 2024, announced the larger mixture-of-experts Rakuten AI 2.0 and a 1.5-billion-parameter mini model that December, and launched the Rakuten AI agentic platform at full scale in July 2025. The service entered Rakuten Link and the web, with a declared goal of becoming the gateway across shopping, lifestyle, fintech, travel, education, work and entertainment.

Rakuten AI 3.0 followed in December 2025 through Japan’s GENIAC program. Its scale and cloud deployment made it a very different tool from the 7B model now placed on HP hardware. Together, the two model families illustrate Rakuten’s architecture: inexpensive and private work close to the user; expensive, connected intelligence in controlled infrastructure; commercial agents above them.

The desktop bundle is distribution as much as technology. Instead of persuading every user to discover a web service, create a habit and pin a tab, Rakuten gains a position in the Windows taskbar and the flow of selected text. HP gains a Japanese service layer that distinguishes its machines from otherwise similar AI PCs. Each company is borrowing the other’s scarce asset: Rakuten supplies local context and services; HP supplies the place where work already happens.

HP’s machine after the cloud revolution

HP’s corporate lineage begins in a Palo Alto garage in 1939, long before the personal computer. The company later became a defining maker of calculators, printers, workstations and PCs, acquired Compaq in 2002, and separated in 2015. HP Inc. retained personal systems and printing; Hewlett Packard Enterprise took enterprise infrastructure, software and services. Rakuten’s partner is the PC company produced by that split.

For much of the cloud era, the personal computer risked becoming a polished terminal: a screen, keyboard and browser through which the most ambitious computation happened elsewhere. The AI PC is the industry’s answer. New machines add a neural processing unit, or NPU, designed to perform large numbers of AI operations efficiently beside the CPU and GPU.

Microsoft formalized one high-end category in May 2024 with Copilot+ PCs. Its requirements include an NPU capable of at least 40 trillion operations per second, 16 gigabytes of memory and 256 gigabytes of storage. HP was among the first hardware partners. Those requirements do not automatically apply to Rakuten AI for Desktop; the Rakuten–HP release defines its own broader eligible-PC bundle and checks local compatibility in the app. The comparison shows what the market now recognizes: useful local AI depends on a hardware floor.

Foundry Local helps abstract that uneven landscape. Microsoft’s June 2026 release described hardware detection and acceleration across Windows AI hardware. The runtime can select an appropriate execution provider instead of making every application team separately optimize for Intel, AMD, Qualcomm or Nvidia components. Rakuten’s model still has to be tuned and tested, but the plumbing is becoming a platform.

HP also points to Wolf Security as the foundation around the workload. That is relevant to device defense; it is not proof that every model output, cloud connection or agent action is safe. A secure boot chain cannot tell whether a generated summary is accurate. Endpoint protection cannot by itself decide whether an agent should see a point balance, read a selected document or confirm a booking.

The route is the real product

In November 2025, Rakuten and HP described “intelligent routing” among cloud, edge and device. The July 2026 product sheet documents a user-accessible switch between on-device and cloud services. It does not publish a decision tree explaining which content is inspected, how any automatic choice is made, whether a task can cross modes midstream or what notice appears before data leaves the machine.

That missing policy matters more than a benchmark score. A local model can summarize a draft contract privately. A cloud model may produce a more capable analysis, but only after receiving content. A shopping agent needs current catalog data. A translation may need neither the internet nor an account. Good routing is not simply “send hard questions to the largest model.” It must account for sensitivity, user consent, latency, network state, price and the permissions needed to act.

Cloud economics supply another reason for the split. Remote inference consumes accelerators every time a user sends and receives tokens. Local inference uses hardware the customer has already purchased, although it still carries electricity, storage, management and support costs. At millions of interactions, moving routine rewrites and summaries to PCs can reduce Rakuten’s marginal bill while giving users quicker responses.

Those incentives do not always align perfectly. The user may prefer a private local answer even when Rakuten could deliver a better cloud result. Rakuten may prefer a connected session that can recommend a product or deepen the customer profile. HP may prefer an experience that demonstrates the value of upgraded hardware. The interface must make the choice legible enough that “hybrid” does not become a euphemism for invisible movement.

Local is more private, not magically private

The privacy case for on-device processing is strong and specific. When local mode handles selected text without a cloud connection, that text need not be transmitted to a remote inference service. Work can continue on an airplane, during an outage or in a location with poor connectivity. Network exposure and remote inference cost are reduced.

But “on-device” is not a blanket security certificate. The PC may still contain malware. Other applications may have clipboard, accessibility or screen-capture permissions. Local prompts and outputs may be logged. Model and software updates arrive through a supply chain. A user can still paste information they are not authorized to process, and a small model can still hallucinate.

Cloud agents add a different class of risk because they connect language to tools and accounts. NIST’s work on agent hijacking includes indirect prompt-injection tests in which malicious content attempts to make an agent exfiltrate files or send phishing messages using the victim’s access. Japan’s AI Guidelines for Business call for privacy protection, security, transparency and risk-based governance across developers, providers and users.

No public evidence at the cutoff suggested Rakuten AI for Desktop had suffered such an attack. The point is architectural: the more an agent can search, book, check balances or navigate subscriptions, the more carefully its permissions and confirmation boundaries must be designed. A fluent answer can be wrong; an agent with excessive authority can make the wrong answer operational.

Enterprises should therefore test local mode rather than accept a privacy adjective. They need to know what the application logs, how administrators control cloud access, where data is processed, how model and agent updates are signed, whether prompts can enter telemetry, how long records remain and whether sensitive categories can be forced to stay offline. HP and Rakuten’s launch materials explain the benefits; procurement requires the controls.

Assistant, agent, merchant

The word “agent” suggests initiative. The documented launch functions are more bounded. Rakuten Ichiba can search and recommend products. Rakuten Travel can search and assist with booking. Point Club can check a balance and answer related questions. The companies describe guidance “from intent to action,” but the public fact sheet does not say an agent may complete an irreversible purchase without explicit user confirmation.

That distinction should remain visible. An assistant helps a person decide. An agent may assemble steps or use a tool. A merchant benefits when the decision ends in a transaction. Rakuten can occupy all three roles inside one interface. Convenience rises because the boundaries disappear; the need for disclosure rises for the same reason.

A recommendation system should tell the user whether results cover the open web, Rakuten inventory or both. Sponsored placement should not masquerade as neutral reasoning. A hotel suggestion should distinguish availability, ranking, loyalty benefits and the commercial interest of the platform. A point-balance query should not grant broader financial visibility than the task needs.

This is the deeper significance of the HP deal. Operating systems once competed to be the desktop. Browsers competed to be the doorway to the web. Search engines competed to answer intention. The agentic layer competes to interpret intention and decide which service gets the next action. A taskbar position on a new PC is valuable because habit can become infrastructure before the user notices a platform forming.

Japan is ready—and still cautious

The timing is favorable. Japan’s 2026 Information and Communications White Paper, released July 24, found that 58.8 percent of surveyed individuals had used generative AI, more than double the preceding survey’s 26.7 percent. The survey expanded to include people aged 15 to 19; among respondents 20 and older, the reported rate was 52.2 percent. China stood at 93.6 percent and the United States and Germany at 75.6 percent.

The increase means AI is no longer a specialist curiosity. The gap means distribution and trust still matter. A Japanese interface already present on a familiar PC can remove the steps of finding a service and deciding which model to use. Offline text functions can address concerns about sending every document to the cloud. Rakuten’s services make the system immediately relevant to ordinary tasks.

Yet bundling can produce installation without meaningful use. Japan’s earlier PC history is full of technically capable products that won or lost through software, compatibility and habit. The local model must be fast enough to feel present, accurate enough to trust and modest enough to admit when a cloud model or human judgment is needed. The agents must save steps without turning every intention into a sales funnel.

The corporate-first release offers a useful proving ground. Businesses can test latency, Japanese quality, deployment, policy controls and offline behavior before the consumer launch from October. It also raises the stakes: enterprise documents are exactly where local processing is most valuable and where unclear logging or mode switching is least tolerable.

The unanswered specification sheet

The launch establishes availability, architecture and functions. It does not yet establish independent performance. Rakuten and HP have not published tokens per second on representative HP models, memory consumption, installation size, battery impact, context length, Japanese summarization accuracy or error rates compared with cloud mode. They have not identified the quantization level used in Rakuten AI 7B ONNX.

The eligible-device wording also needs precision. The bundle covers HP Windows PCs in Japan except gaming PCs, thin clients and some workstations, but the on-device component is for compatible AI PCs. A complete model list, minimum CPU/GPU/NPU and memory requirements, expected performance tiers and support period would let buyers distinguish “comes with an icon” from “runs the local model well.”

Pricing remains open. Official releases do not state a subscription fee. Impress Watch reported from the launch event that the service is free for the time being and that future pricing is under consideration. The eventual division between free local features, metered cloud use, paid agents and benefits tied to Rakuten membership could change the economic meaning of the bundle.

There is also a version paradox. Rakuten’s most capable publicized Japanese model lineage has advanced from 7B to AI 2.0 and AI 3.0, while the PC runs a specially optimized version of the oldest, smallest family. That may be the right engineering choice. Users need a clear update policy: how quality and safety improvements reach the local model, how large downloads are managed and how long a purchased PC remains supported.

Where the sentence goes

In the PC-9801 era, the presence of Japanese on the machine was visible. It appeared in the hardware, the screen modes, the software library and the shape of the market around them. Rakuten AI for Desktop makes the new language layer feel lighter: select a paragraph, touch a contextual bar, receive a rewrite.

Underneath that gesture is a choice with more weight than the old kanji ROM. A sentence processed locally remains part of the user’s machine. A sentence sent to the cloud enters a service relationship. A request given to an agent may become a search, a recommendation, a reservation or another step in a commercial system.

Rakuten and HP have built a credible answer to the engineering problem. A Japanese model can live on the PC; Microsoft’s runtime can help it use heterogeneous hardware; the cloud can supply larger intelligence and current services. The partnership turns “hybrid AI” from an industry slogan into a product people in Japan can activate.

The next achievement must be legibility. Users should know which mind is answering, what it has seen, whose interest it serves and what it is allowed to do. The best hybrid computer will not merely decide where to think. It will make that decision understandable before the thought leaves the desk.

Reporting Notes and Sources

Product, company and market information was checked through July 30, 2026, at 10:12 a.m. JST. Availability refers to the bundle announced for eligible HP Windows PCs in Japan; local mode requires compatible AI hardware as determined by the application. The official launch materials did not provide final pricing, a complete supported-device list, minimum local specifications or independent performance benchmarks. The 58.8-percent adoption figure comes from Japan’s 2026 government white paper survey and is not directly comparable without noting its expanded age range.

• Rakuten and HP: July 29 launch release, availability, eligible-PC exclusions, feature matrix and Rakuten AI 7B ONNX details
• HP Japan: Japanese launch release, activation process, local and cloud modes, security positioning and model background
• Rakuten and HP Japan: November 2025 collaboration and the original intelligent-routing vision
• Impress Watch: July 29 launch-event coverage, current free availability and pricing under consideration
• Microsoft Foundry: Foundry Local 1.2, Windows acceleration and Rakuten AI for Desktop implementation
• Rakuten: March 2024 release of the open Rakuten AI 7B foundation, instruct and chat models
• Rakuten: Rakuten AI 2.0 mixture-of-experts model and 1.5B mini model
• Rakuten: July 2025 full-scale launch of the Rakuten AI agentic platform
• Rakuten: Rakuten AI 3.0 scale, GENIAC support and company-reported Japanese benchmark comparison
• Rakuten: official history from the 1997 marketplace and points program to the Rakuten Ecosystem
• Information Processing Society of Japan Computer Museum: PC-9801 and Japanese-language business computing
• Information Processing Society of Japan Computer Museum: kanji ROM, DOS/V and the development of Japanese PCs
• Microsoft: 2024 introduction of Copilot+ PCs and the 40-plus-TOPS NPU category
• HP: creation of HP Inc. as the personal-systems and printing company in the 2015 separation
• NIST: agent-hijacking evaluations involving prompt injection, data exfiltration and misuse of user permissions
• Ministry of Economy, Trade and Industry: 2026 outline of the AI Guidelines for Business Appendix
• Mainichi / Kyodo: 2026 Information and Communications White Paper adoption results