Ai/Deepseek: Difference between revisions
< Ai
Dascent-wiki (talk | contribs) Created page with "=Deepseek== {{Infobox artificial intelligence | name = DeepSeek AI | logo_filename = File:Deepseek-ai-icon-logo.png|270px|alt=DeepSeek AI logo | developer = DeepSeek (Chinese: 深度求索) | type = Large multimodal model | release_date = Initial release June 2023<br>DeepSeek-R1 (January 2024 | license = Open-source (Apache License 2.0) | website_url = https://www.deepseek.com/ | website_display = Deepseek.com }} '''DeepSeek AI''' is..." |
Dascent-wiki (talk | contribs) No edit summary |
||
| Line 4: | Line 4: | ||
| name = DeepSeek AI | | name = DeepSeek AI | ||
| logo_filename = File:Deepseek-ai-icon-logo.png|270px|alt=DeepSeek AI logo | | logo_filename = File:Deepseek-ai-icon-logo.png|270px|alt=DeepSeek AI logo | ||
| developer = [[DeepSeek]] ( | | developer = [[DeepSeek]] (Cn: 深度求索) | ||
| type = [[Large language model|Large multimodal model]] | | type = [[Large language model|Large multimodal model]] | ||
| release_date = | | release_date = June 2023/DeepSeek-R1 Jan 2024 | ||
| license = Open-source ([[Apache License 2.0]]) | | license = Open-source ([[Apache License 2.0]]) | ||
| website_url = https://www.deepseek.com/ | | website_url = https://www.deepseek.com/ | ||
| website_display = Deepseek.com | | website_display = Deepseek.com | ||
}} | }} | ||
| Line 104: | Line 107: | ||
[[Category:Open-source artificial intelligence]] | [[Category:Open-source artificial intelligence]] | ||
[[Category:2023 software]] | [[Category:2023 software]] | ||
{{Ai/tags|Ai/Deepseek|tg-ai-deepseek}} | |||
Revision as of 11:54, 13 August 2025
Deepseek=
DeepSeek AI
Developer: DeepSeek (Cn: 深度求索)
Type: Large multimodal model
Release Date: June 2023/DeepSeek-R1 Jan 2024
License: Open-source (Apache License 2.0)
Website: Deepseek.com
DeepSeek AI is an open-source artificial intelligence project developed by Chinese AI research company DeepSeek (深度求索). Launched in June 2023, it specializes in developing advanced large language models (LLMs) with particular focus on machine learning research, code generation, and multimodal learning. The project gained significant attention with its January 2024 release of DeepSeek-R1, a 128K-context multimodal model competitive with leading global AI systems.
History
Founding and Early Models (2023)
- June 2023: DeepSeek AI launched with release of DeepSeek-Coder, an open-source code generation model supporting 80+ programming languages
- August 2023: Introduction of DeepSeek-VL, a vision-language multimodal model for image-text understanding
- October 2023: Release of DeepSeek Math, specialized for mathematical reasoning and problem-solving
DeepSeek-R1 Era (2024)
- January 2024: Launch of DeepSeek-R1, featuring:
- 128K token context window
- Multilingual capabilities (English/Chinese focus)
- Strong performance in coding, mathematics, and reasoning benchmarks
- March 2024: Integration with Hugging Face ecosystem and API service launch
- May 2024: Partnership with Linux Foundation for open-source AI infrastructure
Technical Architecture
Model Specifications
| Model | Parameters | Context Window | Specialization | Release Date |
|---|---|---|---|---|
| DeepSeek-Coder | 1.3B-33B | 16K | Code generation | June 2023 |
| DeepSeek-VL | 7B | 32K | Vision-language | August 2023 |
| DeepSeek-R1 | Unknown (est. 30B+) | 128K | General reasoning | January 2024 |
Key Technologies
- Hybrid Attention Mechanism: Combines sliding window attention with global token retention
- Dynamic Token Scaling: Adaptive context management for long documents
- Multimodal Fusion: Cross-modal alignment architecture in DeepSeek-VL
- Code-Centric Pretraining: Specialized datasets with 2:1 code-to-text ratio
Capabilities
- Natural Language Processing: Advanced text generation, summarization, translation
- Programming Assistance: Code completion, debugging, documentation generation
- Mathematical Reasoning: Solving complex equations, theorem proving
- Multimodal Understanding: Image-to-text analysis, visual question answering
- Knowledge Retrieval: Access to current information through web integration
Performance
DeepSeek models consistently rank highly in major AI benchmarks:
- Hugging Face Open LLM Leaderboard: Top 3 open-source models (as of 2024)
- HumanEval: 75.6% pass@1 score (DeepSeek-Coder-33B)
- MMLU: 82.3% accuracy (DeepSeek-R1)
- GSM8K: 86.5% accuracy in mathematical reasoning
Ethical Framework
DeepSeek AI operates under strict ethical guidelines:
- Transparency: Model cards with detailed training data disclosures
- Safety Protocols: Multi-layer content filtering system
- Open Governance: Community input on model deployment policies
- Bias Mitigation: Dedicated adversarial training regimen
Community and Open Source
- GitHub Presence: 15k+ stars across repositories
- Model Hosting: Available on Hugging Face and ModelScope
- Research Partnerships: Collaborations with Tsinghua University and Chinese Academy of Sciences
- License: Open-source under Apache License 2.0 for research/commercial use
Reception
- Praised for "setting new standards in open-source AI" (MIT Technology Review, 2024)
- Recognized as "China's most promising AI research initiative" (South China Morning Post, 2024)
- Critiqued for limited multilingual support beyond Chinese/English
See Also
- Large language model
- Open-source artificial intelligence
- Hugging Face
- Chinese technology
- Artificial intelligence in China
References
References
External Links
This content is generated full or partially by Ai. Click to report inaccurate content.
