DeepSeek V3

📋 Overview

DeepSeek V3 is an open-weight Mixture-of-Experts large language model from DeepSeek AI that delivers GPT-4-class quality at a fraction of the cost. The model uses a 671B total parameter MoE architecture with only 37 billion active parameters per token — achieving exceptional inference efficiency without sacrificing output quality. It is released under the permissive MIT license, allowing unrestricted commercial use, modification, and redistribution of both the model weights and critical evaluation data. DeepSeek V3 is available as open-weight on Hugging Face for self-hosting via vLLM or SGLang, and through DeepSeek's official API and third-party providers like OpenRouter. The model supports 128K context and has demonstrated strong performance across coding, mathematical reasoning, and multilingual tasks, often matching or exceeding GPT-4o on key benchmarks at 1/10th the cost.

✨ Key Features

• •671B MoE architecture with only 37B active params — massive quality at efficient cost
• •MIT licensed — fully open for commercial use, modification, and redistribution
• •Matches GPT-4o quality on many benchmarks at 1/10th the API cost
• •Supports 128K context window for long-form analysis and document processing
• •Strong coding, math, and multilingual capabilities (Chinese, English, and more)
• •Self-hostable on consumer hardware with quantization (GGUF via llama.cpp, Ollama)
• •Available via DeepSeek API, OpenRouter, and HuggingFace Hub
• •OpenAI-compatible API format for easy integration
• •Vibrant community with many fine-tunes, quantizations, and deployment tools