DeepSeek: Search Mode vs. Deep Thinking (R1)

If you're interested in knowing when it's a good idea to DeepSeek Browse the web and when it's worth taking the time to think things through; here you'll find a complete and very practical guide. Throughout the text, you'll see how its two main modes work. the internet search and the deep thinking, in addition to a review of their models, Tricks useful features and ways to use it on mobile, web, or your own PC.

DeepSeek has burst onto the scene as open and free alternative compared to other giants. It not only offers a chatbot accessible to everyone, but also allows you to run models locally and activate a mode of step-by-step reasoning (DeepThink R1) For difficult questions. In these lines we condense everything you need to get the most out of it, with a clear focus on comparing its search and deep thinking modes.

Search Mode vs. Deep Thinking: How They Differently

The overview is simple: when you activate the Search option, you force the IA to consult the web and cite sources, such as the semantic searchWhen you activate Deep Thinking (R1), the system dedicates more computation to breaking down your request and resolving it with multi-step logic. In practice, both are not combined simultaneously. Deep thinking doesn't work in conjunction with online searching.But yes, with documents that you attach.

Search mode is useful for changing content (news, real-time data, recent pages) and, when it works properly, It presents links and updated context.However, it doesn't always get it right; several users have noted that it sometimes fails or doesn't find what's expected, so it's worth checking it.

Deep thinking, on the other hand, uses the R1 reasoning model. This mode mimics a human process: it analyzes, tests hypotheses, and makes decisions step by step. You'll see that, before the final answer, the tool displays the chain of reasoning, often in English (although sometimes it "thinks" in Spanish). It can be activated at any time during the chat. and there is no additional cost in the current official service.

In operational summary: use Search for live, verifiable information on the web; use Deep Thinking when faced with complex challenges in logic, mathematics, programming or analysiswhere the model needs to "stop and think".

What is DeepSeek and why is it on everyone's lips?

DeepSeek is the name of both the Chinese company and the chatbot itself. The firm has chosen to open-source its models, a move that throws off the competition and lowers the barrier to entry for researchers, businesses, and users. Transparency, local usability, and the fact that the chatbot is free. These are their main selling points compared to more closed payment options.

The company works along two main lines: a general-purpose model (V3) and a reasoning model (R1). V3 speaks fluently and writes very well; R1 excels when the difficulty increases and requires fine-tuning with logic, mathematics, or development. Both are available on the app and on the web, and R1 is activated by the Deep Think button.

Compared to other giants, DeepSeek responds directly and to the point, although you can request longer explanations or those tailored to your level. On the economic and operational front, striking figures have been published regarding training costs and token prices, but these need some clarification: Some analysts point out that the actual training figure is not public.So take it as a guideline.

This is how deep thinking works (DeepThink R1)

R1 stems from a very particular approach: large-scale reinforcement learning, with preliminary stages where the model generates and evaluates its own inference traces (R1-Zero). This method enhances its ability to self-analyze reasoning steps and self-correctimproving a lot in areas such as mathematics, logic and coding.

When you activate the deep thinking button, R1 takes more time before returning the final result. During that time, it generates a visible thought process, almost like a conversation with itself, exploring possibilities and discarding errors. The feeling is similar to our "system 2" (thinking slowly) that Kahneman talks about: deliberation before answering.

A notable update to the R1 ecosystem, identified in the documentation as R1-0528, points to performance and stability improvements: greater depth in reasoning, increased tokens available per problem, fewer hallucinations, and better support for structured outputs (JSON and function calls). In public tests, progress has been reported in demanding benchmarks., excelling especially in mathematics and programming.

Very practical points to remember about this mode: you don't need to open a new chat to use it; you can activate it whenever you want and ask them to "think about it again." Furthermore, It works very well with attached documents for in-depth analysis, but it does not combine with internet search.

Internet search mode: when to use it and what to expect

DeepSeek's search function is activated by selecting the Search option just below the text field. At that point, the assistant crawls web content, provides links, and cites recent information. It is especially useful for update data or compare sources when your question cannot be answered with static knowledge of the model.

Even so, it's best to start with realistic expectations: there are cases where the tool doesn't complete the search or doesn't find satisfactory results. In those scenarios, It's a good idea to rethink the question with more context or ask more specific questions, even asking them to summarize first and then elaborate.

A useful trick is to combine phases: start with Search to get links and current overview, then turn off Search and turn on Deep Thinking so that R1 Organize the information, detect contradictions, and propose conclusions more solid on the material already gathered.

Also, remember that online search does not apply when you upload files. If your case involves a PDFa DOCX file or a spreadsheet, R1 can reason directly about the attached content without going online.

V3 or R1: which model to choose for each task

DeepSeek-V3 is your go-to tool for natural conversations, writing, translation, and general language tasks. If fluency and style are what you're after, it's a safe choice. However, when the situation demands chained deductions, logic, or programmingThe R1 model shines because of its structured thought process.

In community and independent lab comparisons, V3 has been seen to compete well with proprietary models in writing, and R1 has matched or surpassed Western benchmarks in mathematics and code in certain tests, and in articles about the best AI for each taskFurthermore, figures have been shared that point to very low operating costs, both in training and tokens. As we mentioned earlier, There are claims that cannot be fully verifiedBut the trend towards efficiency is clear.

If you're unsure, a conservative strategy is to start with V3 for a first draft or basic explanation and then ask R1 to "think" about that output, looking for logical errors, plot holes, or structural improvementsThis combination usually yields very solid results.

Upload files, use OCR, and let AI do the work for you

On mobile and web, you can attach documents. DeepSeek extracts and analyzes their content to answer your questions. With images, current support relies on OCR: it extracts embedded text, but doesn't perform a full analysis. Complete visual understanding of the contentThis could change with the arrival and integration of multimodal models, but for now that is the limitation.

To truly get the most out of it, set clear objectives: "Summarize the document in five key points," "Compare the arguments and suggest improvements," or "Identify potential calculation errors." If you want to raise the bar, activate R1 and request a step-by-step reasoning about the information in the file.

If you work with Excel or CSV, the wizard can help you prepare queries, transform data, and break down results. It's helpful to ask it to explain which formula have you chosen and why?That way you'll understand the criteria and be able to replicate it.

If you work with Excel or CSV, the wizard can help you prepare queries, transform data, and break down results. It's helpful to ask it to explain which formula have you chosen and why?.

Finally, remember that you should take extra precautions when attaching sensitive content. Below you will find privacy recommendations, which are especially relevant in professional environments or with personal data.

Editing, regenerating, and organizing chats

DeepSeek lets you edit your own messages: tap and hold the text (on mobile) or use the chat options to correct a prompt that didn't work. The conversation branches out from that pointto easily compare answers.

You can also regenerate any output. Since the original prompt doesn't change, the result is usually similar, but with variations that sometimes offer improved nuances. Exploring versions is ideal when you're looking for the right tone or structure.

In terms of organization, although folders aren't available today, renaming with emojis works surprisingly well. Visual labels like 📚, 💼, or 🧪 help identify chats related to studying, work, or tests. Additionally, you can delete individual conversations or all of them at once. from settings.

If you ever decide to close your profile, the app offers a direct option to delete your account. The process is simple, although the company may retain certain minimal data according to its policy. It is advisable to review it beforehand. if you are concerned about information retention.

Languages ​​and interface: web and mobile

The web version of DeepSeek is available in English and Chinese, while the mobile app offers more languages, including Spanish. In fact, in bilingual regions, there is support for Basque and Catalan. Android e iOS, and Galician on iOS. Is one of the apps with more languages ​​available that have been seen lately in iPhone.

If you prefer not to install anything, the browser saves space and lets you pin a shortcut to your desktop. All your conversations are synced, so You can switch from mobile to computer without losing track. During peak times, you may see warnings of network congestion or occasional outages.

Four ways to use DeepSeek: app, web, local, and API

The most convenient option for most people is the Android or iOS app: it offers all models and features without a subscription fee. The web service is virtually identical and will get you out of a bind on any device with a browser. Both require a free account and they keep the history; if you want more details about the DeepSeek app on Android You will find specific comparisons.

If privacy is your priority, you can run models locally using tools like Ollama. Install R1 on Windows, macOS (Apple Silicon) or Linux It is feasible (check what can be done with the DeepSeek app on Windows 11The counterpart is clear: There's no internet search and you need good processing power. To enjoy the full performance. Also, local models tend to make more mistakes with specific data if you don't provide it yourself.

To integrate DeepSeek into your own products, there's the official API. Particularly low fees per million reasoning tokens have been mentioned, with costs "starting at $0,014" per million in some ecosystem references. The exact details depend on the provider and the current plan.Platforms like CometAPI offer unified access to multiple models (including deepseek-r1-0528), with competitive pricing and a playground for testing.

Practical tips: Get better answers

Don't settle for the first answer if you're not convinced. Ask for a reformulation, demand that they list assumptions, and that... Check the steps one by one.With R1 active, this persistence translates into more useful chains of reasoning that are less prone to logical errors.

Refine in stages: first request an outline, then the development, then the critical review. In search mode, ask the assistant to distinguish between Facts, opinion, and speculationand that it offers verified links. When you close the search, activate R1 for a more in-depth analysis of the results.

If you program, take advantage of it as a code assistant: generate functions, request unit tests, and ask for complexity upgrades, and compare it with guides on the best AI for programmingHe asks for justification. the choice of structures, patterns, or algorithmsIf an error occurs, paste the stack trace and ask for hypothetical causes and ordered solutions.

To write better, use a reformulator: simplify, shorten, or adapt the tone. If you're studying, turn the content into flashcards or step-by-step explanations. Give concrete examples of what you expect and the AI ​​will follow you better.

Integrations and ecosystem: GitMind and CometAPI

Some third-party tools already integrate DeepSeek's models. GitMind, for example, allows you to choose V3 or R1 for more fluid or reasoning-focused conversations. This facilitates brainstorming, research, and problem-solving sessions. the ideal model for each case.

On the development front, CometAPI unifies access to over 500 models in a single interface. Its catalog includes the deepseek-r1-0528 endpoint, with documentation and playground For testing. If you're looking to move quickly, control costs, and not be tied to a single supplier, this type of multi-model catwalk It can save you a lot of work.

Privacy, censorship, and responsible use

DeepSeek maintains its servers in China. As a precaution, avoid uploading corporate documents, medical results, or other sensitive information if you are not going to run the model locally. In critical environments, the offline route is the safest. because the data never leaves your device.

The free online version may have thematic restrictions consistent with the regulations of the country of origin, especially regarding political and geopolitical issues. This does not affect downloaded and run templates. on your computer, where those restrictions do not applyThis is another reason to consider using local resources if freedom of access is key to your business.

Performance, costs and technical advances

The public narrative surrounding DeepSeek highlights two ideas: high performance and very low costs. Figures have been circulated comparing prices per million tokens and training budgets far below those of other exchanges. Not everything can be verified in detail, but the combination of good performance and open code It is already a fact, and it explains a good part of the impact.

In the recent evolution of reasoning, R1-0528 has added tangible improvements: greater depth of computation per problem, notable reduction of hallucinations in tasks such as summarizing or rewriting, and stronger support for structured output (JSON) and function calls, opening doors to robust integrations.

Work has also been done on distillation towards smaller models (for example, families of 8B parameters) with similar yields in specific mathematical tests. This strategy brings reasoning abilities closer to hardware more modest, with feasible latencies and costs for small teams or embedded deployments.

As historical context, this isn't the first time we've seen a qualitative leap in "thinking before responding." AlphaGo already showed us in 2016 how a system could surprise us with brilliant decisions after thoroughly exploring the game space. R1 brings that idea to language and general problem-solvingdemocratizing it into an accessible and, moreover, open chatbot.

With all of the above in mind, the final picture is clear: if you want live, traceable information, use Search mode; if you need solid arguments and logical chains of reasoning, activate Deep Thinking. Alternate between both as the task requires. Rely on archives, message editing, and regenerations to refine the result until it meets your needs.