Skip to content
TheChatArchiveThe Chat Archive
Perplexitysonar-pro-high· PerplexityHugging Face LMArena Search Arena 24K

please compare grok-3 and deepseek R1

From a Chatbot Arena battleThe same person sent this prompt to two anonymous models at once. Compare it with gpt-4o-search-preview-high’s answer.
  1. User

    please compare grok-3 and deepseek R1

  2. Perplexity Assistant

    Here's a comparison of Grok-3 and DeepSeek R1 based on the available information:

    Architecture and Training

    Grok-3:

    • Built on major infrastructure upgrades and new training techniques[13]
    • Trained on xAI's Colossus supercluster with 10x the compute of previous state-of-the-art models[13]
    • Uses reinforcement learning to refine its chain-of-thought process[13]

    DeepSeek R1:

    • Developed using a combination of reinforcement learning and supervised fine-tuning[14]
    • Incorporates 671 billion parameters, with 37 billion active at any given time[2]
    • Uses a Mixture of Experts (MoE) architecture[2]

    Performance and Capabilities

    Grok-3:

    • Achieved 93.3% on the 2025 American Invitational Mathematics Examination (AIME)[13]
    • Scored 84.6% on graduate-level expert reasoning (GPQA)[13]
    • Attained 79.4% on LiveCodeBench for code generation and problem-solving[13]
    • Has a context window of 1 million tokens[13]

    DeepSeek R1:

    • Scored 79.8% on AIME 2024, slightly ahead of OpenAI o1-1217 at 79.2%[14]
    • Achieved 97.3% on MATH-500, surpassing OpenAI o1-1217 at 96.4%[14]
    • Scored 96.3% on Codeforces, close to OpenAI o1-1217's 96.6%[14]
    • Performed strongly on SWE-bench Verified with 49.2%, slightly ahead of OpenAI o1-1217's 48.9%[14]

    Specialized Features

    Grok-3:

    • Includes DeepSearch, a built-in search engine for real-time information[1]
    • Has a "Big Brain" mode for complex problem-solving[1]
    • Can process text, code, and images[1]

    DeepSeek R1:

    • Optimized for English and Chinese, but may have limitations with other languages[14]
    • Can be fine-tuned for specific tasks or industries[14]
    • Processes text and potentially images and audio[19]

    Accessibility and Deployment

    Grok-3:

    • Currently in beta, being rolled out to users[13]
    • Accessible through xAI's platform[13]

    DeepSeek R1:

    • Open-source and available under an MIT license[14]
    • Can be run locally with high-performance GPUs or TPUs[14]
    • Available as a fully managed serverless model in Amazon Bedrock[18]

    Cost and Efficiency

    Grok-3:

    • Specific cost information not provided in the search results

    DeepSeek R1:

    • Training costs were under 6million,comparedtoanestimated6 million, compared to an estimated 100 million for GPT-4[5]
    • More cost-efficient due to its MoE architecture[2]

    In summary, both Grok-3 and DeepSeek R1 are advanced AI models with strong performance in reasoning, mathematics, and coding tasks. Grok-3 appears to have a slight edge in some benchmarks and offers unique features like DeepSearch and Big Brain mode. DeepSeek R1, on the other hand, stands out for its open-source nature, cost-efficiency, and flexibility in deployment options. The choice between the two would depend on specific use cases, required features, and deployment preferences.

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.