DeepSeek-R1-0528 API

Reasoning models are trained very differently from their non-reasoning counter parts, and as a result they serve different purposes. Below we'll compare both types of models, details for reasoning models, pros and cons, applications and example use-cases.

Reasoning models like DeepSeek-R1 are specifically developed to engage in extended, deep analysis of complex challenges. Their strength lies in strategic thinking, developing comprehensive solutions to intricate problems, and processing large amounts of nuanced information to reach decisions. Their high precision and accuracy make them particularly valuable in specialized fields traditionally requiring human expertise, such as mathematics, scientific research, legal work, healthcare, financial analysis.

Non-reasoning models such as Llama 3.3 70B or DeepSeek-V3 are trained for efficient, direct task execution with faster response times and better cost efficiency.

Your application can leverage both types of models: using DeepSeek-R1 to develop the strategic framework and problem-solving approach, while deploying non-reasoning models to handle specific tasks where swift execution and cost considerations outweigh the need for absolute precision.

Reasoning models excel for tasks where you need:

• High accuracy and dependable decision-making capabilities
• Solutions to complex problems involving multiple variables and ambiguous data
• Can afford higher query latencies
• Have a higher cost/token budget per task

Non-reasoning models are optimal when you need:

• Faster processing speed(lower overall query latency) and lower operational costs
• Execution of clearly defined, straightforward tasks
• Function calling, JSON mode or other well structured tasks

1. Introduction

The DeepSeek R1 model has undergone a minor version upgrade, with the current version being DeepSeek-R1-0528. In the latest update, DeepSeek R1 has significantly improved its depth of reasoning and inference capabilities by leveraging increased computational resources and introducing algorithmic optimization mechanisms during post-training. The model has demonstrated outstanding performance across various benchmark evaluations, including mathematics, programming, and general logic. Its overall performance is now approaching that of leading models, such as O3 and Gemini 2.5 Pro.

Compared to the previous version, the upgraded model shows significant improvements in handling complex reasoning tasks. For instance, in the AIME 2025 test, the model’s accuracy has increased from 70% in the previous version to 87.5% in the current version. This advancement stems from enhanced thinking depth during the reasoning process: in the AIME test set, the previous model used an average of 12K tokens per question, whereas the new version averages 23K tokens per question.

Beyond its improved reasoning capabilities, this version also offers a reduced hallucination rate, enhanced support for function calling, and better experience for vibe coding.

2. Evaluation Results

DeepSeek-R1-0528

For all our models, the maximum generation length is set to 64K tokens. For benchmarks requiring sampling, we use a temperature of $0.6$, a top-p value of $0.95$, and generate 16 responses per query to estimate pass@1.

Note: We use Agentless framework to evaluate model performance on SWE-Verified. We only evaluate text-only prompts in HLE testsets.  GPT-4.1 is employed to act user role in Tau-bench evaluation.

DeepSeek-R1-0528-Qwen3-8B

Meanwhile, we distilled the chain-of-thought from DeepSeek-R1-0528 to post-train Qwen3 8B Base, obtaining DeepSeek-R1-0528-Qwen3-8B. This model achieves state-of-the-art (SOTA) performance among open-source models on the AIME 2024, surpassing Qwen3 8B by +10.0% and matching the performance of Qwen3-235B-thinking. We believe that the chain-of-thought from DeepSeek-R1-0528 will hold significant importance for both academic research on reasoning models and industrial development focused on small-scale models.

3. How to Run Locally

Please visit DeepSeek-R1 repository for more information about running DeepSeek-R1-0528 locally.

Compared to previous versions of DeepSeek-R1, the usage recommendations for DeepSeek-R1-0528 have the following changes:

1. System prompt is supported now.
2. It is not required to add "<think>\n" at the beginning of the output to force the model into thinking pattern.

The model architecture of DeepSeek-R1-0528-Qwen3-8B is identical to that of Qwen3-8B, but it shares the same tokenizer configuration as DeepSeek-R1-0528. This model can be run in the same manner as Qwen3-8B.

System Prompt

In the official DeepSeek web/app, we use the same system prompt with a specific date.

For example,

Temperature

In our web and application environments, the temperature parameter $T_{model}$ is set to 0.6.

Prompts for File Uploading and Web Search

For file uploading, please follow the template to create prompts, where {file_name}, {file_content} and {question} are arguments.

For Web Search, {search_results}, {cur_date}, and {question} are arguments.For Chinese query, we use the prompt:

For English query, we use the prompt:

5. License

This code repository is licensed under MIT License. The use of DeepSeek-R1 models is also subject to MIT License. DeepSeek-R1 series (including Base and Chat) supports commercial use and distillation.

6. Citation

Prompting DeepSeek-R1, and other reasoning models in general, is quite different from working with non-reasoning models.

Below we provide guidance on how to get the most out of DeepSeek-R1:

• Clear and specific prompts: Write your instructions in plain language, clearly stating what you want. Complex, lengthy prompts often lead to less effective results.
• Sampling parameters: Set the temperature within the range of 0.5-0.7 (0.6 is recommended) to prevent endless repetitions or incoherent outputs. Also, a top-p of 0.95 is recommended.
• No system prompt: Avoid adding a system prompt; all instructions should be contained within the user prompt.
• No few-shot prompting: Do not provide examples in the prompt, as this consistently degrades model performance. Rather, describe in detail the problem, task, and output format you want the model to accomplish. If you do want to provide examples, ensure that they align very closely with your prompt instructions.
• Structure your prompt: Break up different parts of your prompt using clear markers like XML tags, markdown formatting, or labeled sections. This organization helps ensure the model correctly interprets and addresses each component of your request.
• Set clear requirements: When your request has specific limitations or criteria, state them explicitly (like "Each line should take no more than 5 seconds to say..."). Whether it's budget constraints, time limits, or particular formats, clearly outline these parameters to guide the model's response.
• Clearly describe output: Paint a clear picture of your desired outcome. Describe the specific characteristics or qualities that would make the response exactly what you need, allowing the model to work toward meeting those criteria.
• Majority voting for responses: When evaluating model performance, it is recommended to generate multiple solutions and then use the most frequent results.
• No chain-of-thought prompting: Since these models always reason prior to answering the question, it is not necessary to tell them to "Reason step by step..."
• Math tasks: For mathematical problems, it is advisable to include a directive in your prompt such as: "Please reason step by step, and put your final answer within \boxed{}."
• Forcing <think>: On rare occasions, DeepSeek-R1 tends to bypass the thinking pattern, which can adversely affect the model's performance. In this case, the response will not start with a <think> tag. If you see this problem, try telling the model to start with the <think> tag.

Reasoning models use-cases

• Analyzing and assessing AI model outputs: Reasoning models excel at evaluating responses from other systems, particularly in data validation scenarios. This becomes especially valuable in critical fields like law, where these models can apply contextual understanding rather than just following rigid validation rules.

• Code analysis and improvement: Reasoning models are great at conducting thorough code reviews and suggesting improvements across large codebases. Their ability to process extensive code makes them particularly valuable for comprehensive review processes.

• Strategic planning and task delegation: These models shine in creating detailed, multi-stage plans and determining the most suitable AI model for each phase based on specific requirements like processing speed or analytical depth needed for the task.

• Complex document analysis and pattern recognition: The models excel at processing and analyzing extensive, unstructured documents such as contract agreements, legal reports, and healthcare documentation. They're particularly good at identifying connections between different documents and making connections.

• Precision information extraction: When dealing with large volumes of unstructured data, these models excel at pinpointing and extracting exactly the relevant information needed to answer specific queries, effectively filtering out noise in search and retrieval processes. This makes them great to use in RAG or LLM augmented internet search use-cases.

• Handling unclear instructions: These models are particularly skilled at working with incomplete or ambiguous information. They can effectively interpret user intent and will proactively seek clarification rather than making assumptions when faced with information gaps.

Reasoning models are trained very differently from their non-reasoning counter parts, and as a result they serve different purposes. Below we'll compare both types of models, details for reasoning models, pros and cons, applications and example use-cases.

Reasoning models like DeepSeek-R1 are specifically developed to engage in extended, deep analysis of complex challenges. Their strength lies in strategic thinking, developing comprehensive solutions to intricate problems, and processing large amounts of nuanced information to reach decisions. Their high precision and accuracy make them particularly valuable in specialized fields traditionally requiring human expertise, such as mathematics, scientific research, legal work, healthcare, financial analysis.

Non-reasoning models such as Llama 3.3 70B or DeepSeek-V3 are trained for efficient, direct task execution with faster response times and better cost efficiency.

Your application can leverage both types of models: using DeepSeek-R1 to develop the strategic framework and problem-solving approach, while deploying non-reasoning models to handle specific tasks where swift execution and cost considerations outweigh the need for absolute precision.

Reasoning models excel for tasks where you need:

• High accuracy and dependable decision-making capabilities
• Solutions to complex problems involving multiple variables and ambiguous data
• Can afford higher query latencies
• Have a higher cost/token budget per task

Non-reasoning models are optimal when you need:

• Faster processing speed(lower overall query latency) and lower operational costs
• Execution of clearly defined, straightforward tasks
• Function calling, JSON mode or other well structured tasks