DeepSeek

DeepSeek is a Chinese AI company founded in 2023, dedicated to making AGI a reality. Based in Hangzhou, DeepSeek is known for producing frontier-level open-weight models at a fraction of typical training costs.

Notable Models

DeepSeek-V3 Series — Flagship general-purpose models (685B MoE)

• DeepSeek-V3.2 — Reasoning-first model built for agents (December 2025)
• DeepSeek-V3.1 — General-purpose flagship
• DeepSeek-V3.1-Terminus — Optimized variant

DeepSeek-R1 Series — Reasoning models trained via reinforcement learning (January 2025)

• DeepSeek-R1 — 671B MoE (37B active); on par with OpenAI o1
• DeepSeek-R1-Distill — Distilled variants based on Qwen and Llama (1.5B to 70B)

Specialized Models

• DeepSeek-Coder-V2 — Code generation
• DeepSeek-Math-V2 — Mathematical reasoning
• DeepSeek-VL2 — Vision-language model
• DeepSeek-OCR — Optical character recognition (3B)

Multimodal

• Janus-Pro-7B — Unified multimodal understanding and generation

Key Innovations

• Mixture of Experts (MoE): 685B total parameters with only 37B active per forward pass
• Reinforcement Learning for Reasoning: R1 trained via large-scale RL without supervised fine-tuning
• Cost Efficiency: V3 reportedly trained for ~$6M vs. $100M+ for comparable models

Links

• Website
• DeepSeek Chat
• API Platform
• API Documentation
• HuggingFace
• GitHub
• X (Twitter)