AI・テクノロジー
DeepSeek R2 Shock | The Cheapest, Most Powerful Reasoning AI That Runs on a Single RTX GPU
機械翻訳 / Machine-translated
A
AI Friends
2026年4月26日・24分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
"DeepSeek has shaken the world again" — In April 2026, Chinese AI company DeepSeek released a new reasoning AI model called "R2."
With a relatively compact size of 32B (32 billion) parameters, it achieved 92.7% on the challenging math benchmark "AIME 2025."
It delivers performance nearly on par with OpenAI o3, while API pricing is 30–70% or less of major competitors — a true price disruption.
And it's fully open-sourced under the MIT license.
This article provides a thorough breakdown of what DeepSeek R2 is and what it means for developers and businesses in Japan.
DeepSeek R2 is a reasoning-specialized AI model released in April 2026 by Chinese AI company "DeepSeek." A "reasoning AI" is not the kind that simply outputs an answer — it works through a "thinking process," much like solving a math problem step by step, to arrive at a precise answer.
Think of it as the difference between someone who does mental math quickly and someone who carefully works through it with pencil and paper.
Reasoning AI is the "pencil-and-paper type" — fewer mistakes, but it takes a little more time.
It particularly excels in situations where errors are unacceptable, such as math, coding, and logic puzzles.
DeepSeek originally made waves in January 2025 when it released a reasoning AI called "R1," which offered performance surpassing GPT-4 at dramatically low prices.
R2 is its successor.
Perhaps the most intuitive way to describe it: "faster than a sports car, but runs on the fuel budget of a compact car."
The biggest surprise with DeepSeek R2 is that it "got smarter while getting smaller."
DeepSeek R1, released in January 2025, was an enormous MoE (Mixture of Experts — a system that combines multiple specialized AI models) model with a staggering 671 billion (671B) parameters.
While highly capable, running it yourself required a GPU cluster worth millions of yen or more.
It was like a giant machine that needed an entire professional baseball team to operate.
R2, by contrast, adopts a "dense transformer" architecture with 32 billion (32B) parameters.
Rather than MoE, it uses a straightforward design where all parameters are utilized.
It is roughly 1/21st the size of R1, yet its scores on challenging benchmarks are equal to or better than R1's.
It runs on a single consumer gaming GPU, the "NVIDIA RTX 4090" (24GB).
Imagine handpicking just the "ace with a .400 batting average" from an entire professional baseball team — and getting the same game results.
This achievement of "higher performance at smaller size" was made possible by a reinforcement learning method called GRPO (Group Relative Policy Optimization) and "distillation."
Specifically, the massive R1 was used as the "teacher model" to generate millions of "thinking traces" — the step-by-step reasoning processes used to solve math, coding, and logic problems.
The 32B R2 (the "student model") then learned from these, efficiently inheriting the teacher's knowledge — like a student receiving all the lecture notes from an outstanding teacher.
Let's answer the question: "Just how impressive is 92.7%?"
AIME (American Invitational Mathematics Examination) is an extremely difficult test that serves as the qualifying round for the United States Mathematical Olympiad.
At the highest level of high school mathematics, scoring 92.7% out of 15 questions means getting approximately 14 questions correct.
Even top-tier high school students in the US are considered exceptional if they answer 8–10 correctly.
Comparison with major models:
R2 represents an improvement of roughly 18 points or more over R1 — a remarkable leap for a single generation.
Even at equivalent performance, the pricing is dramatically different. API pricing comparison (per 1 million input tokens):
If you were spending ¥100,000 per month on a reasoning AI, switching alone could bring that down to roughly ¥30,000 or less. The impact is like discovering that a premium convenience store coffee tastes exactly the same as a supermarket store brand.
There are two main reasons DeepSeek R2 is this affordable.
The MIT license is the most permissive open-source license, allowing free commercial use, modification, and redistribution.
In other words, you can download the R2 model files for free and run them on your own server.
Zero API costs are possible.
In music terms, it's like using a royalty-free track directly in a commercial product.
While competitors like OpenAI and Anthropic keep their models closed, DeepSeek is pursuing an open-source strategy to draw in developers worldwide.
DeepSeek R2's API is fully compatible with OpenAI's API format.
This means existing ChatGPT-integrated systems can be switched to DeepSeek R2 simply by changing the "API URL" and "API key."
In Python code, it's just two lines: base_url and api_key.
Like moving into a new place and not having to buy new appliances — just change the plug and everything works.
There are three main ways to use DeepSeek R2 in Japan. Choose based on your goals and budget.
Simply obtain an API key from DeepSeek's official website and connect to the OpenAI-compatible endpoint.
You can use a GPT-4-class reasoning AI for a cost ranging from a few thousand to tens of thousands of yen per month.
Even without programming skills, it can be easily integrated using no-code tools or existing frameworks like LangChain.
Download the model files from Hugging Face and run them entirely locally on a PC with an RTX 4090 (24GB VRAM).
Zero API costs, and since data never leaves your machine, it's reassuring from a security standpoint.
It means the era has arrived where you can have a top-tier reasoning AI on your own server — well-suited for fields that handle sensitive data, such as finance, healthcare, and law.
DeepSeek models are also becoming available on Microsoft Azure and AWS.
If you can run them in a Japan domestic region, it becomes easier to meet compliance requirements that prohibit data from leaving the country.
This is especially advantageous for companies in finance, healthcare, and the public sector, and lets you leverage your existing cloud infrastructure as-is.
What kind of impact will DeepSeek R2 have in Japan? Let's explore three concrete scenarios.
E-san develops AI chatbots at a Tokyo startup.
Previously using OpenAI o3, the company was paying over ¥500,000 per month in API fees.
Switching to DeepSeek R2 could bring that down to under ¥150,000 per month with equivalent performance.
The freed-up ¥350,000 can be reinvested into feature development.
This is the kind of case where "reasoning AI adoption that was too expensive to consider" becomes a reality — a particularly revolutionary change for startups and independent developers.
F-san's firm wanted to introduce an AI that could read through case law and summarize legal opinions.
However, there were concerns about sending clients' confidential information to servers in the United States.
Running R2 locally means data stays entirely within the firm's own servers.
It's reassuring from a legal ethics standpoint, and a PC with an RTX 4090 costs around ¥1,000,000 — which could be lower in long-term cost than paying monthly API fees.
G-san wants to use AI to analyze factory defect inspection data, but faces the constraint that data containing equipment manufacturers' proprietary know-how cannot be sent to an external cloud.
Running R2 on-premises (on the company's own equipment) allows high-precision reasoning AI to be used while keeping confidential data protected.
R2 offers one answer to a problem Japanese companies have long struggled with: balancing confidentiality obligations with AI adoption.
On the other hand, since DeepSeek is a Chinese company, some have raised concerns about data handling.
In particular, when using the official API, prompts are sent to servers in China, so caution is warranted for use cases involving confidential information.
These concerns can be resolved by running R2 locally or self-hosting the open-source version on your own servers.
Since it is open-source under the MIT license, you can review the code yourself — a level of transparency that closed models cannot offer.
A. Yes, it supports Japanese.
It is capable enough for practical use in tasks like business email drafting, translation, and coding assistance.
However, there are cases where Japanese accuracy falls slightly short of English, so for critical use cases, it is recommended to try prompting in English.
A. For use cases where reasoning performance matters — math, coding, complex logic problems — it is worth considering a switch.
R2 has improved over R1 by approximately 18 points or more on AIME, making it clearly superior in terms of performance.
On the other hand, for simple text generation or conversation, general-purpose models like DeepSeek V3.2 may offer better cost-efficiency.
Choose based on your use case.
A. If you are running it locally or self-hosting it on a cloud server, you can continue operating even if DeepSeek goes out of business.
Because the model is publicly released under the MIT license, copies also exist on platforms like Hugging Face.
However, if you are only using the official API, there is a risk of service discontinuation.
For critical business systems, consider "self-hosting."
A. For local execution, a GPU with 24GB VRAM or more is recommended, but you can use it via the official API or cloud services from any PC or smartphone.
Using quantization (a technique to compress the model), it may also be possible to run it on GPUs with 16GB or 8GB (with some reduction in performance).
Starting with the official API is recommended.
A. R2 is positioned as a "reasoning specialist (Reasoner)," while V3.2 is a "general-purpose conversation and code generation" model.
R2 delivers high accuracy on math, logic, and coding, but processing is somewhat slower.
The basic guideline is: use R2 for "problems that require a definitive answer," and V3.2 for "everyday work where you want fast and reasonably accurate responses."
While GPT-5 and Claude 4 are being developed with investments in the trillions of yen, DeepSeek continues blazing its own trail of "maximum performance from minimal resources."
The arrival of R2 is an event that shatters the conventional wisdom that "high-performance reasoning AI is only for large enterprises and research institutions."
We recommend starting by creating an account with DeepSeek's official API and trying it out on a small project.
This article is a cross-post from AI Friends.
