How to use the new DeepSeek R1 safely

DeepSeek R1 is the latest release in AI (LLMs), emerging as a strong competitor to OpenAI's O1 model. With superior performance in various natural language processing tasks, it has been attracting attention from both experts and companies seeking more advanced Artificial Intelligence solutions. In this post, we'll explore why DeepSeek R1 is trending, how it stands out in terms of performance, and most importantly, how to use it safely through Inner AI.

One of the central factors explaining DeepSeek R1's rapid rise is its ability to deliver superior results compared to OpenAI's O1 model while requiring only 3% of the resources that the competing solution needs. This data not only draws attention for its expressive efficiency gain but also highlights the innovative nature of the model's architecture, based on advanced Reinforcement Learning techniques. In this approach, DeepSeek R1 is not limited to processing large volumes of text; it uses continuous evaluation and adjustment mechanisms that allow it to "reason" iteratively.

Beyond the direct impact on costs and performance, DeepSeek R1's sustainable approach has also gained prominence. By demanding less computational infrastructure, it reduces energy consumption and, consequently, the carbon footprint — an increasingly valued point in a market seeking to balance innovation and environmental responsibility. All this efficiency strongly impacted the international financial scene: shortly after the announcement of DeepSeek R1's success, AI sector companies like Nvidia, Meta, and Alphabet (Google) faced significant drops in their stocks. The fear of a possible "disruption" driven by Chinese technology led some investors to review their positions, reinforcing the perception that DeepSeek R1 inaugurates a new phase in the language model market.

DeepSeek R1 Performance in AI Benchmarks

To consolidate its status as a heavyweight competitor to OpenAI's O1, DeepSeek R1 underwent various formal tests. Inner AI closely monitors these results to offer the best user experience to its clients. Below, we see how the model performed in Mathematics, Programming, and General Knowledge areas.

Mathematics Benchmarks: AIME 2024 and MATH-500

One of the great challenges in language models is the ability to handle complex mathematical problems, which require not only textual interpretation but also logical reasoning and application of advanced concepts. In the AIME 2024 tests (famous for evaluating mathematical skills at a competitive level) and MATH-500, which brings together an extensive set of algebra, geometry, and combinatorial analysis questions, DeepSeek R1 demonstrated high precision in problem-solving and step-by-step explanation of solutions.

Surpassing O1: According to internal comparisons, R1 showed greater consistency in handling complex algebraic expressions and formal proofs, indicating that its Reinforcement Learning architecture allows refining each step of reasoning.
Lower Error Margin: At the end of the test batteries, DeepSeek R1 presented a slightly lower error rate than O1, suggesting greater ability to "correct its own course" as it progresses in solving each question.

This performance reinforces the idea that the model not only understands natural language in its statements but also uses internal mechanisms to validate mathematical solutions before presenting them.

Coding Benchmarks: Codeforces & SWE-bench Verified

In an increasingly software-oriented world, the ability to understand and generate code is a significant differential for AI models. In Codeforces, a programming competition platform with thousands of problems in various languages, and in SWE-bench Verified, which evaluates generated code quality and functionality, DeepSeek R1 maintained consistent results and, in some cases, surpassed O1.

Reading More Complex Problems: When analyzing questions with advanced algorithms or intricate data structures, R1 displayed greater clarity and lower incidence of "logical bugs" in the resulting code.
Syntax and Best Practices: Besides producing functional code, the model followed style conventions closer to those adopted by the development community, facilitating readability and maintenance of created snippets.

The agility with which DeepSeek R1 suggests solutions also drew attention. Even when faced with extensive problems, it maintained a competitive response time, a factor that makes it particularly attractive for prototyping applications and development team support.

General Knowledge Benchmarks: GPQA Diamond and MMLU

The ability to answer general questions, ranging from current events to history, is evaluated in benchmarks such as GPQA Diamond (which covers general knowledge questions and contextual reasoning) and MMLU, focused on diverse areas such as humanities and exact sciences. In these tests, DeepSeek R1 proved especially proficient in:

Deepening Context: When questioned about historical events, scientific concepts, or general culture, R1 presented more detailed answers than O1, demonstrating extensive vocabulary and greater ability to relate facts.
Multiple Domains: The model managed to navigate well through distinct topics, going from literature questions to chemistry problems, without significant loss of precision.

This versatile performance reinforces R1's proposal to be a comprehensive solution for various AI demands, allowing it to be applied both in teaching platforms and customer service services.

Security and privacy with Inner AI

With DeepSeek R1's popularization, concerns arose around data security, especially due to original terms that allow the use of customer information for continuous improvement purposes. Inner AI, however, developed a robust infrastructure that neutralizes these risks, ensuring all DeepSeek R1 users a safe and transparent way to leverage the model's potential.

Local Storage Instead of routing data to Chinese servers, Inner AI maintains all infrastructure in data centers in Brazil, USA, and Europe. Thus, besides compliance with international data protection legislation, we ensure that client information remains within trusted regions, free from transfer to China or any other unauthorized country.
Compliance with Official Policies In parallel, we strictly follow the security and privacy guidelines declared by DeepSeek itself, as described in the documents at DeepSeekR1.org. However, we reinforce our protection layers to block any unauthorized data collection attempts.
Protection against Data Reuse Even though DeepSeek's original policy mentions the reuse of inputs and outputs for training, Inner AI does not engage in this practice. Each interaction is isolated and encrypted, so we don't contribute client information to model training. Thus, the confidentiality and secrecy of your operations are assured.
Advanced Security Protocols and SOC2 Compliance All standards and processes adopted by Inner AI — from end-to-end encryption to constant threat monitoring — can be consulted on our portal trust.innerai.com. There, you'll find details about regular audits and certifications that ensure data inviolability, including compliance with the SOC2 (Service Organization Control 2) standard. This certification, validated by independent audits, attests that we follow a rigorous set of controls to protect client information privacy and security.

See how to use DeepSeek R1 on Inner AI

1. Create or access your Inner AI account If you're not yet a user, register here on Inner AI's official page. If you already have an account, simply log in with your credentials.

2. Access the Inner AI panel After entering the platform, on the home page, in the upper left corner, you'll find the icon with available model options. Click on "AI Models," then scroll down to "Deep Reasoning" and select DeepSeek R1 or DeepSeek R1 Small.

3. Define the context or prompt To start using the model, enter in the text field the prompt or description of the problem you want to solve. It's at this stage that you provide the necessary information for the AI to understand what needs to be generated or answered.

Inner AI is proud to be the first Brazilian company to make DeepSeek R1 available with full Portuguese support, ensuring simplified and secure access through its platform. Our infrastructure is fully protected, with no training or sharing of your information with third parties, including the Chinese model manufacturer.