DeepSeek R1 vs OpenAI o1: Which One is Faster, Cheaper and Smarter?

Pankaj Singh Last Updated : 29 Jan, 2025

DeepSeek R1它已经到来，它不仅仅是另一种人工智能模型，它是人工智能能力的重大飞跃，是在之前发布的[DeepSeek-V3-Base变体]上训练的(https://www.analyticsvidhya.com/blog/2024/12/deepseek-v3/). 随着DeepSeek R1的全面发布，它现在在性能和灵活性方面与OpenAI o1不相上下。更引人注目的是它的开放权重和MIT许可，使其在商业上可行，并将其定位为开发人员和企业的有力选择。

The [DeepSeek R1]has arrived, and it’s not just another AI model—it’s a significant leap in AI capabilities, trained upon the previously released DeepSeek-V3-Base variant. With the full-fledged release of DeepSeek R1, it now stands on par with OpenAI o1 in both performance and flexibility. What makes it even more compelling is its open weight and MIT licensing, making it commercially viable and positioning it as a strong choice for developers and enterprises alike.

但真正让DeepSeek R1与众不同的是它如何挑战像OpenAI这样的行业巨头，用一小部分资源取得显著成果。在短短两个月内，DeepSeek完成了看似不可能的事情——推出了一个与专有系统竞争的开源人工智能模型，同时在严格的限制下运行。在本文中，我们将比较DeepSeek R1和OpenAI o1。

But what truly sets DeepSeek R1 apart is how it challenges industry giants like OpenAI, achieving remarkable results with a fraction of the resources. In just two months, DeepSeek has done what seemed impossible—launching an open-source AI model that rivals proprietary systems, all while operating under strict limitations. In this article, we will compare – DeepSeek R1 vs OpenAI o1.

Table of contents

1. DeepSeek R1: A Testament to Ingenuity and Efficiency
2. What Makes DeepSeek R1 a Game-Changer?
3. Overview of DeepSeek R1
4. How DeepSeek R1 Gives Unbeatable Performance at Minimal Cost?
5. DeepSeek R1 vs. OpenAI o1: Price Comparison
6. DeepSeek R1 vs OpenAI o1: Comparison of Different Benchmarks
7. How to Access DeepSeek R1 Using Ollama?
8. How to Use DeepSeek R1 in Google Colab?
9. Code Implementation of OpenAI o1
10. Conclusion

1.[DeepSeek R1:独创性和效率的见证]-deepseek-r1-a测试真实性和效率
2.[是什么让DeepSeek R1成为游戏规则的改变者？]
3.【DeepSeek R1概述】
4.[DeepSeek R1如何以最低成本提供无与伦比的性能？]
5.[DeepSeek R1与OpenAI o1：价格比较]
6.[DeepSeek R1与OpenAI o1：不同基准的比较]
7.[如何使用Olama访问DeepSeek R1？]
8.[如何在谷歌Colab中使用DeepSeek R1？]
9.【OpenAI o1的代码实现】
10.[结论]

DeepSeek R1: A Testament to Ingenuity and Efficiency

With a budget of just $6 million , DeepSeek has accomplished what companies with billion-dollar investments have struggled to do. Here’s how they did it:

• Budget Efficiency: Built R1 for just $5.58 million , compared to OpenAI’s estimated $6 billion+ investment.
• Resource Optimization: Achieved results with 2.78 million GPU hours , significantly lower than Meta’s 30.8 million GPU hours for similar-scale models.
• Innovative Workarounds: Trained using restricted Chinese GPUs , showcasing ingenuity under technological and geopolitical constraints.
• Benchmark Excellence: R1 matches OpenAI o1 in key tasks, with some areas of clear outperformance.

While DeepSeek R1 builds upon the collective work of open-source research, its efficiency and performance demonstrate how creativity and strategic resource allocation can rival the massive budgets of Big Tech.

DeepSeek的预算仅为600万美元，它完成了拥有数十亿美元投资的公司一直在努力做的事情。以下是他们是如何做到的：
预算效率： 构建R1仅需 558万美元，而OpenAI的估计投资为60多亿美元
资源优化： 实现了278万GPU小时的结果，明显低于Meta的**3080万GPU小时，适用于类似规模的模型。
创新解决方案： 使用 受限的中国GPU进行培训，展示在技术和地缘政治限制下的独创性。
卓越基准测试： R1在关键任务上与OpenAI o1相匹配，在某些领域表现明显优于OpenAI

虽然DeepSeek R1建立在开源研究的集体工作之上，但其效率和性能表明创造力和战略资源分配可以与大型科技公司的巨额预算相媲美。

What Makes DeepSeek R1 a Game-Changer? 是什么让DeepSeek R1成为游戏规则的改变者？

Beyond its impressive technical capabilities, DeepSeek R1 offers key features that make it a top choice for businesses and developers:

• Open Weights & MIT License: Fully open and commercially usable, giving businesses the flexibility to build without licensing constraints.
• Distilled Models: Smaller, fine-tuned versions (akin to Qwen and Llama), providing exceptional performance while maintaining efficiency for diverse applications.
• API Access: Easily accessible via API or directly on their platform—for free!
• Cost-Effectiveness: A fraction of the cost compared to other leading AI models, making advanced AI more accessible than ever.

DeepSeek R1 raises an exciting question—are we witnessing the dawn of a new AI era where small teams with big ideas can disrupt the industry and outperform billion-dollar giants? As the AI landscape evolves, DeepSeek’s success highlights that innovation, efficiency, and adaptability can be just as powerful as sheer financial might.

除了令人印象深刻的技术能力外，DeepSeek R1还提供了使其成为企业和开发人员首选的关键功能：
开放权重和MIT许可证： 完全开放且可商业使用，使企业能够在没有许可限制的情况下灵活构建。
蒸馏型号： 更小、微调的版本（类似于Qwen和Llama），在为各种应用保持效率的同时提供卓越的性能。
API访问： 通过API或直接在其平台上轻松访问-免费
成本效益： 与其他领先的人工智能模型相比，成本只是其中的一小部分，使先进的人工智能比以往任何时候都更容易获得。
DeepSeek R1提出了一个令人兴奋的问题——我们是否正在见证一个新的人工智能时代的黎明，在这个时代，有大想法的小团队可以颠覆行业，超越数十亿美元的巨头？随着人工智能领域的发展，DeepSeek的成功凸显了创新、效率和适应性与纯粹的财务实力一样强大。

Overview of DeepSeek R1 DeepSeek R1概述

DeepSeek R1模型拥有6710亿个参数架构，并在DeepSeek V3 Base模型上进行了训练。它对思维链（CoT）推理的关注使其成为需要高级理解和推理的任务的有力竞争者。有趣的是，尽管其参数数量很大，但在大多数操作中只有370亿个参数被激活，类似于DeepSeek V3。
DeepSeek R1不仅仅是一个整体模型；该生态系统包括六个蒸馏模型，这些模型基于DeepSeek R1本身的合成数据进行了微调。这些较小的模型在大小和目标特定用例方面各不相同，为需要更轻、更快模型的开发人员提供了解决方案，同时保持了令人印象深刻的性能。

The DeepSeek R1 model boasts a 671 billion parameters architecture and has been trained on the DeepSeek V3 Base model. Its focus on Chain of Thought (CoT) reasoning makes it a strong contender for tasks requiring advanced comprehension and reasoning. Interestingly, despite its large parameter count, only 37 billion parameters are activated during most operations, similar to DeepSeek V3.

DeepSeek R1 isn’t just a monolithic model; the ecosystem includes six distilled models fine-tuned on synthetic data derived from DeepSeek R1 itself. These smaller models vary in size and target specific use cases, offering solutions for developers who need lighter, faster models while maintaining impressive performance.

Distilled Model Lineup

Model	Base Model	Download
DeepSeek-R1-Distill-Qwen-1.5B	Qwen2.5-Math-1.5B	🤗 HuggingFace
DeepSeek-R1-Distill-Qwen-7B	Qwen2.5-Math-7B	🤗 HuggingFace
DeepSeek-R1-Distill-Llama-8B	Llama-3.1-8B	🤗 HuggingFace
DeepSeek-R1-Distill-Qwen-14B	Qwen2.5-14B	🤗 HuggingFace
DeepSeek-R1-Distill-Qwen-32B	Qwen2.5-32B	🤗 HuggingFace
DeepSeek-R1-Distill-Llama-70B	Llama-3.3-70B-Instruct	🤗 HuggingFace

这些经过提炼的模型实现了灵活性，同时满足了本地部署和API的使用。值得注意的是，Llama 33.7B型号在几个基准测试中表现优于o1 Mini，突显了蒸馏变体的优势。

These distilled models enable flexibility, catering to both local deployment and API usage. Notably, the Llama 33.7B model outperforms the o1 Mini in several benchmarks, underlining the strength of the distilled variants.

Model	#Total Params	#Activated Params	Context Length	Download
DeepSeek-R1-Zero	671B	37B	128K	🤗 HuggingFace
DeepSeek-R1	671B	37B	128K	🤗 HuggingFace

You can find all about OpenAI o1 here.

How DeepSeek R1 Gives Unbeatable Performance at Minimal Cost? DeepSeek R1如何以最低的成本提供无与伦比的性能？

DeepSeek R1以最低的成本取得了令人印象深刻的性能，这得益于其培训和优化过程中的几个关键策略和创新。以下是他们是如何做到的：

DeepSeek R1’s impressive performance at minimal cost can be attributed to several key strategies and innovations in its training and optimization processes. Here’s how they achieved it :

1. Reinforcement Learning Instead of Heavy Supervised Fine-Tuning 强化学习代替繁重的监督微调

Most traditional LLMs (like GPT, LLaMA, etc.) rely heavily on supervised fine-tuning, which requires extensive labeled datasets curated by human annotators. DeepSeek R1 took a different approach :

• DeepSeek-R1-Zero:
  • Instead of supervised learning, it utilized pure reinforcement learning (RL) .
  • The model was trained through self-evolution , allowing it to iteratively improve reasoning capabilities without human intervention.
  • RL helps in optimizing policies based on trial-and-error, making the model more cost-effective compared to supervised training, which requires vast human-labeled datasets.
• DeepSeek-R1 (Cold Start Strategy):
  • To avoid common issues in RL-only models (like incoherent responses), they introduced a small, high-quality supervised dataset for a “cold start.”
  • This enabled the model to bootstrap better from the beginning, ensuring human-like fluency and readability while maintaining strong reasoning capabilities.

Impact:

• RL training significantly reduced data annotation costs.
• Self-evolution allowed the model to discover problem-solving strategies autonomously.

大多数传统的LLM（如GPT、LLaMA等）严重依赖于监督微调，这需要由人类注释者管理的大量标记数据集。DeepSeek R1采取了不同的方法：

DeepSeek-R1-Zero：
*它使用了纯强化学习（RL），而不是监督学习。
*该模型通过自我进化进行训练，使其能够在没有人为干预的情况下迭代地提高推理能力。
*RL有助于基于试错法优化策略，与需要大量人类标记数据集的监督训练相比，使模型更具成本效益。
DeepSeek-R1（冷启动策略）：
*为了避免仅RL模型中的常见问题（如不连贯的响应），他们引入了一个小型、高质量的监督数据集用于“冷启动”
*这使得模型从一开始就能够更好地引导，确保人类般的流畅性和可读性，同时保持强大的推理能力。
影响：
*RL训练显著降低了数据注释成本。
*自我进化使模型能够自主发现解决问题的策略

2. Distillation for Efficiency and Scaling

Another game-changing approach used by DeepSeek was the distillation of reasoning capabilities from the larger R1 models into smaller models, such as:

• Qwen, Llama, etc.
  • By distilling knowledge, they were able to create smaller models (e.g., 14B) that outperform even some state-of-the-art (SOTA) models like QwQ-32B.
  • This process essentially transferred high-level reasoning capabilities to smaller architectures, making them highly efficient without sacrificing much accuracy.

Key Distillation Benefits:

• Lower computational costs: Smaller models require less inference time and memory.
• Scalability: Deploying distilled models on edge devices or cost-sensitive cloud environments is easier.
• Maintaining strong performance: The distilled versions of R1 still rank competitively in benchmarks.

2.蒸馏以提高效率和结垢
DeepSeek使用的另一种改变游戏规则的方法是将推理能力从较大的R1模型提炼为较小的模型，例如：
*Qwen、Llama等
*通过提取知识，他们能够创建更小的模型（例如14B），其性能甚至优于QwQ-32B等一些最先进的（SOTA）模型。
*这个过程基本上将高级推理能力转移到了较小的架构中，使它们在不牺牲太多准确性的情况下高效运行。
蒸馏的主要好处：
***较低的计算成本：**较小的模型需要较少的推理时间和内存。
***可扩展性：**在边缘设备或成本敏感的云环境上部署蒸馏模型更容易。
***保持强劲的性能：**R1的蒸馏版本在基准测试中仍然具有竞争力。

3. Benchmark Performance & Optimization Focus

DeepSeek R1 has focused its optimization towards specific high-impact benchmarks like:

• AIME 2024: Achieving near SOTA performance at 79.8%
• MATH-500: Improving reasoning with 97.3% accuracy
• Codeforces (Competitive Programming): Ranking within the top 3.7%
• MMLU (General Knowledge): Competitive at 90.8%, slightly behind some models, but still impressive.

Instead of being a general-purpose chatbot, DeepSeek R1 focuses more on mathematical and logical reasoning tasks, ensuring better resource allocation and model efficiency.

3.基准性能和优化重点
DeepSeek R1将优化重点放在特定的高影响力基准上，如：
AIME 2024: 实现接近SOTA的性能，达到79.8%
MATH-500: 提高推理准确率97.3%
Codeforce（竞争性编程）： 排名前3.7%
MMLU（一般知识）： 竞争力为90.8%，略落后于一些型号，但仍然令人印象深刻。
DeepSeek R1不是一个通用的聊天机器人，而是更专注于数学和逻辑推理任务，确保更好的资源分配和模型效率。

4. Efficient Architecture and Training Techniques

DeepSeek likely benefits from several architectural and training optimizations:

• Sparse Attention Mechanisms:
  • Enables processing of longer contexts with lower computational cost.
• Mixture of Experts (MoE):
  • Possibly used to activate only parts of the model dynamically, leading to efficient inference.
• Efficient Training Pipelines:
  • Training on well-curated, domain-specific datasets without excessive noise.
  • Use of synthetic data for reinforcement learning phases.

4.高效的架构和训练技术
DeepSeek可能从几个架构和训练优化中受益：
*注意力分散机制：
*能够以较低的计算成本处理较长的上下文。
*混合专家（MoE）：
*可能仅用于动态激活模型的部分，从而实现高效推理。
*高效的培训管道：
*在精心策划的、特定领域的数据集上进行培训，而不会产生过多噪音。
*在强化学习阶段使用合成数据。

5. Strategic Model Design Choices

DeepSeek’s approach is highly strategic in balancing cost and performance by:

1. Focused domain expertise (math, code, reasoning) rather than general-purpose NLP tasks.
2. Optimized resource utilization to prioritize reasoning tasks over less critical NLP capabilities.
3. Smart trade-offs like using RL where it works best and minimal fine-tuning where necessary.

5.战略模型设计选择
DeepSeek的方法在平衡成本和性能方面具有高度的战略性：
1.专注于领域专业知识（数学、代码、推理），而不是通用的NLP任务。
2.优化资源利用率，将推理任务优先于不太关键的NLP能力。
3.明智的权衡比如在RL最有效的地方使用RL，在必要时进行最小的微调。

Why Is It Cost-Effective?

• Reduced need for expensive supervised datasets due to reinforcement learning.
• Efficient distillation ensures top-tier reasoning performance in smaller models.
• Targeted training focus on reasoning benchmarks rather than general NLP tasks.
• Optimization of architecture for better compute efficiency.

By combining reinforcement learning, selective fine-tuning, and strategic distillation , DeepSeek R1 delivers top-tier performance while maintaining a significantly lower cost compared to other SOTA models.

####为什么它具有成本效益？

*由于强化学习，减少了对昂贵的监督数据集的需求。
*高效的蒸馏确保了在较小模型中的顶级推理性能。
*有针对性的训练侧重于推理基准，而不是一般的NLP任务。
*优化架构以提高计算效率。

通过结合强化学习、选择性微调和战略蒸馏，DeepSeek R1提供了顶级性能，同时与其他SOTA型号相比保持了显著更低的成本。

DeepSeek R1 vs. OpenAI o1: Price Comparison DeepSeek R1与OpenAI o1：价格比较

Source: DeepSeek

DeepSeek R1 scores comparably to OpenAI o1 in most evaluations and even outshines it in specific cases. This high level of performance is complemented by accessibility; DeepSeek R1 is free to use on the DeepSeek chat platform and offers affordable API pricing. Here’s a cost comparison:

• DeepSeek R1 API : 55 Cents for input, $2.19 for output ( 1 million tokens)
• OpenAI o1 API : $15 for input, $60 for output ( 1 million tokens)

API is 96.4% cheaper than chatgpt.

DeepSeek R1’s lower costs and free chat platform access make it an attractive option for budget-conscious developers and enterprises looking for scalable AI solutions.

DeepSeek R1在大多数评估中的得分与OpenAI o1相当，在特定情况下甚至超过了它。这种高水平的性能得到了可访问性的补充；DeepSeek R1可在DeepSeeke聊天平台上免费使用，并提供负担得起的API定价。以下是成本比较：

*DeepSeek R1 API：输入55美分，输出2.19美元（100万代币）
*OpenAI o1 API：输入15美元，输出60美元（100万代币）

API比chatgpt便宜96.4%

DeepSeek R1的低成本和免费聊天平台访问使其成为注重预算的开发人员和寻求可扩展人工智能解决方案的企业的一个有吸引力的选择。

Benchmarking and Reliability 基准测试和可靠性

DeepSeek models have consistently demonstrated reliable benchmarking, and the R1 model upholds this reputation. DeepSeek R1 is well-positioned as a rival to OpenAI o1 and other leading models with proven performance metrics and strong alignment with chat preferences. The distilled models, like Qwen 32B and Llama 33.7B , also deliver impressive benchmarks, outperforming competitors in similar-size categories.

DeepSeek模型一直表现出可靠的基准测试，R1模型也维护了这一声誉。DeepSeek R1是OpenAI o1和其他领先模型的竞争对手，具有经过验证的性能指标，并与聊天偏好高度一致。经过提炼的型号，如Qwen 32B和Llama 33.7B，也提供了令人印象深刻的基准，在类似尺寸类别中表现优于竞争对手。

Practical Usage and Accessibility 实用性和可访问性

DeepSeek R1 and its distilled variants are readily available through multiple platforms:

1. DeepSeek Chat Platform : Free access to the main model.
2. API Access : Affordable pricing for large-scale deployments.
3. Local Deployment : Smaller models like Qwen 8B or Qwen 32B can be used locally via VM setups.

While some models, such as the Llama variants, are yet to appear on AMA, they are expected to be available soon, further expanding deployment options.

DeepSeek R1及其蒸馏变体可通过多个平台轻松获得：

1.DeepSeek聊天平台：免费访问主模型。
2.API访问：大规模部署的价格合理。
3.本地部署：较小的型号，如Qwen 8B或Qwen 32B，可以通过VM设置在本地使用。

虽然一些型号，如Llama变种，尚未出现在AMA上，但预计很快就会上市，进一步扩大了部署选项。

DeepSeek R1 vs OpenAI o1: Comparison of Different Benchmarks 不同基准的比较

Source: DeepSeek

1. AIME 2024 (Pass@1)

• DeepSeek-R1: 79.8% accuracy
• OpenAI o1-1217: 79.2% accuracy
• Explanation:
  • This benchmark evaluates performance on the American Invitational Mathematics Examination (AIME), a challenging math contest.
  • DeepSeek-R1 slightly outperforms OpenAI-o1-1217 by 0.6%, meaning it’s marginally better at solving these types of math problems.
    *该基准评估了美国邀请数学考试（AIME）的表现，这是一项具有挑战性的数学竞赛。
    DeepSeek-R1的表现略优于OpenAI-o1-1217 0.6%， 这意味着它在解决这些类型的数学问题方面略胜一筹。

2. Codeforces (Percentile)

• DeepSeek-R1: 96.3%
• OpenAI o1-1217: 96.6%
• Explanation:
  • Codeforces is a popular competitive programming platform, and percentile ranking shows how well the models perform compared to others.
  • OpenAI-o1-1217 is slightly better (by 0.3%), meaning it may have a slight advantage in handling algorithmic