DeepSeek-V3
By DeepSeek
DeepSeek-V3 is a large open-weight mixture-of-experts language model released by DeepSeek in December 2024, serving as the general-purpose foundation model that the reasoning-focused DeepSeek-R1 was later built on.
Definition
DeepSeek-V3 is a large open-weight mixture-of-experts language model released by DeepSeek in December 2024, serving as the general-purpose foundation model that the reasoning-focused DeepSeek-R1 was later built on.
Overview
DeepSeek-V3 is a large-scale mixture-of-experts (MoE) model with a very large total parameter count, of which only a small fraction is activated for any given token, following the efficiency-focused MoE approach popularized by models such as Mixtral 8x7B. DeepSeek reported strong performance across general knowledge, coding, and math benchmarks at release, competitive with leading contemporaries, while emphasizing training efficiency and lower reported compute cost relative to comparable dense models from other labs. As a general-purpose foundation model, DeepSeek-V3 was designed for broad chat, coding, and instruction-following use, distinct from the specialized step-by-step reasoning behavior later added in DeepSeek-R1, which was trained on top of the V3 base using additional reinforcement learning. DeepSeek released V3's weights openly, contributing to its rapid adoption for self-hosting and further fine-tuning, and cementing DeepSeek's place alongside Qwen and Llama as one of the most prominent open-weight model families competing with proprietary frontier labs.
Key Features
- Large-scale mixture-of-experts architecture for training and inference efficiency
- General-purpose foundation model for chat, coding, and instruction-following
- Base model underlying the DeepSeek-R1 reasoning model
- Openly released weights enabling self-hosting and fine-tuning
- Competitive benchmark performance reported at a lower training cost
- Released by DeepSeek in December 2024