DeepSeek V3 - The latest open-source AI model from DeepSeek, with programming capabilities surpassing Claude.

What is DeepSeek V3?

DeepSeek V3 is the latest open-source AI model from DeepSeek, the AI company under the well-known private equity giant, Magic Square Quant. It has made significant progress in multi-language programming capabilities. Its performance in the Aider multi-language programming benchmark surpasses competitors such as Claude 3.5 and Sonnet V2. The latest version, DeepSeek-V3-0324, boasts 685 billion parameters, employs a hybrid expert architecture, and integrates multi-head latent attention and multi-label prediction technologies, achieving an inference speed of 20 tokens per second. In terms of programming assistance, DeepSeek-V3-0324 performs exceptionally well, easily generating complete front-end pages and replacing most daily coding tasks for junior programmers. It possesses powerful long text processing capabilities, capable of untangling complex logical chains, making it suitable for academic research and document analysis. Its dialogue interaction capabilities have also been improved, resulting in more natural and fluent conversations. DeepSeek V3 is now open-source and can be viewed on Hugging Face.

Main functions of DeepSeek V3

• Code generationIt can quickly generate high-quality code based on natural language descriptions input by users, and supports multiple programming languages such as Python, JavaScript, and Java.
• Code optimization: Analyze and optimize existing code to improve its performance, readability, and maintainability.
• Debugging assistanceIt helps users find errors and vulnerabilities in their code, providing detailed error information and repair suggestions. When users encounter code execution errors, they can quickly locate the problem and find possible solutions, just like an experienced programmer.
• Text generationIt can generate long, coherent texts, such as novels, papers, and reports, based on topics or outlines provided by users.
• Text AnalysisThis involves in-depth analysis of long texts, extracting key information, summarizing main content, and outlining logical structure.
• Text polishingAdjusting and polishing the language style of user-provided text to make it more fluent, natural, and in line with a specific writing style.
• Natural DialogueIt can engage in fluent and natural conversations with users, understand their questions, and provide accurate answers.
• Multiple rounds of dialogueIt supports multi-turn dialogues and can understand and answer subsequent questions based on contextual information, maintaining the coherence and consistency of the dialogue.
• Personalized dialogueProvide a personalized conversational experience based on user preferences and history.
• Training efficiencySupports FP8 mixed-precision training, improving training speed and reducing GPU memory usage. Designs the DualPipe algorithm to achieve efficient pipelined parallel processing. Optimizes cross-node All-to-All communication, fully utilizing InfiniBand and NVLink bandwidth.
• Pre-training and post-trainingPre-training was performed on 14.8T tokens, and the context window was expanded from 4K to 128K through a two-stage context expansion. Supervised fine-tuning and reinforcement learning were then performed to align with human preferences and further unlock potential.
• Performance evaluationDeepSeek-V3's base model performs exceptionally well in multiple standard and open benchmark tests, particularly in code and mathematics. The chat version of DeepSeek-V3 also outperforms other open-source models and rivals the performance of leading closed-source models.
• Cost-effectivenessThe training cost is only 2.788M H800 GPU hours, and the total cost is 5.576M USD.
• APIs and Web ServicesDeepSeek provides APIs and web services to facilitate integration and use in different scenarios.
• Multilingual processing capabilitiesDeepSeek V3 has made a major breakthrough in multi-language programming capabilities, outperforming competitors such as Claude 3.5 and Sonnet V2 in the aider multi-language programming assessment.

DeepSeek V3 Technical Principles

• Architecture DesignDeepSeek V3 employs a Mixture-of-Experts (MoE) architecture, comprising up to 256 experts, each a separate neural network capable of handling specific tasks or data types. In the MoE architecture, not all experts participate in every computation; instead, a selection of experts is dynamically chosen through a routing mechanism (such as sigmoid routing). In DeepSeek V3, each computation selects the top 8 most relevant experts (topk=8).
• Working mechanismIt is divided into the following key stages:
  • planBased on user queries, plan the format of the final results and define the entity types and related columns to be extracted.
  • searchCombining keyword search and neural search, and with the support of Exa, it accurately locates content.
  • extract: Utilize large language models (LLM) to efficiently identify and extract specific information from content.
  • RichFurther populate the extracted data to ensure that each entry is detailed and complete.
• Multimodal capabilitiesDeepSeek V3 uses OCRvl2 technology, which can better preserve text, formatting, and formulas in images, achieving results that surpass traditional OCR.
• Streaming rendering optimizationThe web page uses streaming output, but since the Markdown needs to be re-parsed for each rendering, the current rendering speed of 60tps may cause some delay.

DeepSeek V3 project address

• HuggingFace model library:
  • https://huggingface.co/collections/deepseek-ai/deepseek-v3
  • https://huggingface.co/deepseek-ai/DeepSeek-V3-0324
• Paper link:https://github.com/deepseek-ai/DeepSeek-V3/blob/main/DeepSeek_V3.pdf
• Product ExperienceDeepSeek

Performance and efficiency improvements in DeepSeek V3

• Parameter sizeDeepSeek V3 employs a MoE architecture with up to 685 billion parameters. This large-scale parameterization enables the model to capture more complex patterns and relationships.
• Computing resource managementThrough the MoE architecture, DeepSeek V3 can dynamically select the most suitable expert for computation, thereby reducing unnecessary computation and memory consumption.
• Data parallelism and model parallelismDeepSeek V3 employs parallel strategies such as data parallelism, tensor parallelism, sequence parallelism, and 1F1B pipeline parallelism during training. These strategies improve hardware utilization and accelerate model training.
• Optimized learning rate schedulerDeepSeek V3 uses a multi-stage learning rate scheduler, which helps the model maintain the optimal learning rate at different training stages.
• Scaling Laws ResearchThe DeepSeek V3 development team conducted in-depth research on Scaling Laws to find the optimal model/data size allocation ratio and to predict the training results of large-scale models.
• Security assessmentDeepSeek V3 performs rigorous data security screening throughout the entire training process to ensure that the trained model conforms to human values.

Multiple benchmark scores for DeepSeek V3

• In the authoritative test set HumanEvalDeepSeek-V3-0324 scored 328.3 points, surpassing Claude 3.7's 322.3 points.
• In LiveBench testingDeepSeek V3 scored very high, indicating that it can quickly respond to user queries and provide feedback.
  • Global average score60.4 points
  • reasoning ability50 points
  • Programming skills63.4 points
  • Mathematical Analysis60 points
  • Data Analysis57.7 points
  • Language comprehension50.2 points
  • Immediate Feedback (IF)80.9 points
• Educational benchmark tests:
  • In the MMLU (Multiple-Topic Multiple-Choice) benchmark, DeepSeek-V3 achieved an accuracy of 88.5%, surpassing all other open-source models and comparable to leading closed-source models such as GPT-4o and Claude-Sonnet-3.5.
  • In MMLU-Pro (a more rigorous multi-topic multiple-choice test), DeepSeek-V3 achieved an accuracy of 75.9%, which also outperformed other open-source models and was comparable to the performance of top closed-source models.
• Factual benchmarking:
  • In SimpleQA and Chinese SimpleQA, DeepSeek-V3 outperformed GPT-4o and Claude-Sonnet-3.5 in terms of Chinese factual knowledge, demonstrating its advantage in this area.
• Code, math, and reasoning benchmarks:
  • DeepSeek-V3 outperforms all open-source and closed-source non-long-CoT models in math-related benchmarks, even surpassing o1-preview, such as MATH-500, in some benchmarks, demonstrating its powerful mathematical reasoning capabilities.
  • In programming-related tasks, DeepSeek-V3 became the best-performing model in LiveCodeBench (real-time code benchmarking), solidifying its leading position in this field.
• Open assessment:
  • In open-ended dialogue evaluations, DeepSeek-V3 achieved excellent results in the AlpacaEval 2.0 and Arena-Hard benchmark tests, demonstrating its powerful ability to handle complex prompts and tasks.
• Evaluation as a generative reward model:
  • In the RewardBench evaluation, DeepSeek-V3 showed performance comparable to GPT-4o and Claude-3.5, further demonstrating its ability as a reward model.

DeepSeek V3 API service pricing

• Pricing:Every million inputs Tokens are 0.5 yuan (cache hit) / 2 yuan (cache miss), per million outputs 8 tokens Yuan.
• Promotional trial periodFrom now until 2025 Year 2 moon 8 On that day, a 45-day trial period with discounted prices was set.
  • Promotional Price: Per million inputs tokens 0.1 (cache hit) / 1 Yuan (cache miss), per million outputs tokens 2 Yuan.
  • rights and interestsExisting users and new users who register during this period can enjoy discounted prices.

Application scenarios of DeepSeek V3

• Education and TrainingDeepSeek V3 can serve as a personalized learning assistant, providing instant answers and guidance based on students' learning progress and circumstances.
• Content creationWriters and content creators can use DeepSeek V3 to get inspiration, generate content outlines, or polish text.
• Scientific explorationResearchers can use DeepSeek V3 for complex scientific research tasks such as data analysis and pattern recognition.
• Product DevelopmentThrough the DeepSeek API, developers can seamlessly integrate the AI capabilities of DeepSeek V3 into various applications and products, enhancing their intelligence.
• Information retrievalDeepSeek V3 aims to revolutionize the existing search engine concept, transforming it from a simple "answer engine" into a more powerful "search engine".